Photovoltaic hydrogen production power feed-forward control method and system based on shadow space-time prediction

Abstract

The invention relates to the technical field of power control, in particular to a photovoltaic hydrogen production power feed-forward control method and system based on shadow space-time prediction. The method comprises the following steps: constructing a spatio-temporal data acquisition system, acquiring solar irradiance, atmospheric temperature, cloud layer motion trail and component surface shadow distribution data of a photovoltaic array area through a multi-dimensional sensor network, dividing a photovoltaic array into a plurality of monitoring units, and performing spatio-temporal alignment processing on real-time data, and forming a standardized data set comprising the spatial position, the timestamp and the multi-dimensional physical quantity. By constructing a multi-dimensional spatio-temporal data acquisition system, high-precision real-time sensing of photovoltaic array area shadow distribution and environmental parameters is realized, a prediction architecture fused with spatio-temporal characteristics is combined, spatial correlation between monitoring units and a time sequence rule of shadow migration are effectively captured, and prospective input is provided for power control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power control, in particular to a photovoltaic hydrogen production power feedforward control method and system based on shadow spatiotemporal prediction. BACKGROUND

[0002] As an important path for efficient utilization of renewable energy, the stability and energy conversion efficiency of photovoltaic hydrogen production are significantly affected by solar radiation conditions. Photovoltaic arrays are easily disturbed by local shadows caused by factors such as cloud movement and vegetation obstruction during operation, leading to nonlinear fluctuations in power generation. The dynamic response capability of hydrogen production equipment to input power has threshold constraints. Traditional control strategies rely on real-time feedback adjustment and are difficult to predict power fluctuation trends in advance during the spatiotemporal coupling process of shadow changes, resulting in frequent adjustment of equipment operating parameters, reduced energy conversion efficiency, and even safety risks.

[0003] In the prior art, there is a lack of systematic modeling of shadow migration laws in the spatiotemporal scale, and the control strategy fails to effectively integrate prediction information and equipment operating characteristics to form a closed-loop control system with foresight and adaptability. Therefore, the technical problem addressed by the present application is how to construct a collaborative control architecture covering shadow spatiotemporal feature collection, high-precision prediction, and power dynamic matching to improve the stability of photovoltaic hydrogen production systems. SUMMARY

[0004] The present disclosure proposes a photovoltaic hydrogen production power feedforward control method and system based on shadow spatiotemporal prediction, aiming to overcome at least one of the deficiencies in the prior art.

[0005] To achieve the above-mentioned purpose, the technical solutions disclosed by the present application are as follows: According to one aspect of the present disclosure, a photovoltaic hydrogen production power feedforward control method based on shadow spatiotemporal prediction is provided, the steps of the method comprising: Constructing a spatiotemporal data collection system, acquiring solar irradiance, atmospheric temperature, cloud movement trajectory, and component surface shadow distribution data of the photovoltaic array region through a multi-dimensional sensor network, dividing the photovoltaic array into multiple monitoring units and performing spatiotemporal alignment processing on real-time data to form a standardized data set including spatial position, timestamp, and multi-dimensional physical quantities; Establishing a shadow spatiotemporal prediction model, processing the standardized data set based on a prediction architecture that integrates spatiotemporal features, extracting spatial neighborhood shadow features and time series migration laws of the monitoring units, and generating shadow coverage probability and intensity prediction values for each monitoring unit within a specified future time; Developing a power feedforward control strategy, calculating the photovoltaic array power fluctuation range based on the prediction values, generating feedforward control instructions in combination with the power response characteristics of the hydrogen production equipment, and controlling the fluctuation amplitude within the safe operating range of the equipment by adjusting the input power parameters of the electrolytic cell. A dynamic feedback correction mechanism is constructed to collect the actual power generation of the photovoltaic array and the operating parameters of the hydrogen production equipment in real time, calculate the deviation between the actual power generation and the predicted and control target parameters, and dynamically correct the predicted model parameters and control strategy compensation parameters when the deviation exceeds the preset threshold, thus forming a closed-loop optimization control system.

[0006] Furthermore, the multi-dimensional sensor network includes a spectroradiometer, an infrared thermal imager, and a dynamic vision sensor deployed in the photovoltaic array area. The monitoring unit division is based on the principle of spatial continuity, with each monitoring unit corresponding to at least three cross-validation data acquisition points from the sensors. The standardized dataset also includes photovoltaic module type, installation tilt angle, and historical shading event records.

[0007] Furthermore, the prediction architecture that integrates spatiotemporal features includes a spatial feature extraction unit and a time series analysis unit. The spatial feature extraction unit uses a hierarchical convolution method to process the neighborhood shadow features of the monitoring unit. The time series analysis unit captures the temporal correlation of shadow migration through a gated loop mechanism. The output layer of the shadow spatiotemporal prediction model simultaneously outputs the confidence index of shadow prediction.

[0008] Furthermore, the power feedforward control strategy includes a power compensation threshold determination module, which generates a power adjustment amplitude coefficient based on the shadow coverage probability, shadow intensity, and equipment load rate. The electrolytic cell input power adjustment follows the energy conservation constraint and the equipment rated power boundary condition, and adopts a segmented adjustment method to control the power change rate.

[0009] Furthermore, the deviation calculation of the dynamic feedback correction mechanism includes multi-dimensional analysis of power deviation and equipment status parameter deviation. The dynamic correction process combines equipment operating status feedback data and historical correction records, and selectively adjusts the prediction model and control strategy parameters through a preset optimization algorithm. The optimization algorithm sets a correction frequency limit to avoid system oscillation. The equipment operating status feedback data includes the electrolyte temperature, pressure, and conductivity of the electrolyzer. The historical correction records include correction events and parameter adjustment records within a preset time period. The dynamic correction process differentiates the parameter correction weights by establishing an equipment health model.

[0010] Furthermore, the spatiotemporal data acquisition system also accesses data from meteorological satellites and ground meteorological stations. Through a data fusion algorithm, macroscopic meteorological information is combined with local sensor data. The data fusion algorithm uses a state estimation framework to process external variables such as cloud movement speed and solar incidence angle, thereby improving the predictive model's ability to respond to regional weather changes.

[0011] Furthermore, the method also includes a pre-planning stage, which establishes a typical scenario library based on historical data, including different weather types, shadow durations and coverage areas, and pre-generates control parameter combinations for each scenario. When the real-time prediction result matches a scenario in the scenario library, the corresponding pre-planned control parameters are directly invoked.

[0012] Furthermore, the power compensation threshold determination module integrates a logic judgment unit. The input variables include the shadow prediction results and equipment operating status parameters, and the output variables are the adjustment instructions for the electrolytic cell's operating voltage and current. The generation process of the adjustment instructions takes into account the equipment's thermal response characteristics and power mutation constraints.

[0013] Furthermore, the process of generating the shadow coverage probability and intensity prediction values ​​includes weighting the spatial location correlation of the monitoring units. The weighting is based on the distance between the monitoring units and the frequency of historical shadow co-occurrence, so as to strengthen the analysis of the mutual influence of shadow features in adjacent areas.

[0014] According to another aspect of this disclosure, a photovoltaic hydrogen production power feedforward control system based on shading spatiotemporal prediction is provided for implementing the photovoltaic hydrogen production power feedforward control method based on shading spatiotemporal prediction as described above, the system comprising: The data acquisition and preprocessing module acquires spatiotemporal data of the photovoltaic array area through a multi-dimensional sensor network, which is used to divide the monitoring units and perform spatiotemporal alignment processing to generate a standardized dataset. The shadow spatiotemporal prediction module processes the standardized dataset based on a prediction architecture that integrates spatiotemporal features, and is used to generate shadow coverage probability and intensity prediction values ​​for each monitoring unit. The power feedforward control module calculates the power fluctuation range based on the predicted value and generates feedforward control commands to adjust the input power parameters of the electrolytic cell to control the fluctuation amplitude. The dynamic feedback correction module collects actual operating parameters and calculates deviations. When the deviation exceeds a preset threshold, it dynamically corrects the prediction model and control strategy parameters to form a closed-loop optimization control system. The data acquisition and preprocessing module accesses external meteorological data and performs fusion processing; the shadow spatiotemporal prediction module includes a spatial feature extraction unit and a time series analysis unit; the power feedforward control module sets a power compensation threshold determination and segmented adjustment mechanism; and the dynamic feedback correction module integrates a multi-dimensional deviation analysis unit and a parameter selective adjustment unit.

[0015] The beneficial effects of this invention are: This invention constructs a multi-dimensional spatiotemporal data acquisition system to achieve high-precision real-time perception of the shadow distribution and environmental parameters of the photovoltaic array area. Combined with a predictive architecture that integrates spatiotemporal features, it effectively captures the spatial correlation between monitoring units and the time series pattern of shadow migration, providing forward-looking input for power control.

[0016] The power feedforward control strategy proposed in this invention deeply couples the shading prediction results with the power response characteristics of the hydrogen production equipment. By dynamically adjusting the input parameters of the electrolyzer, it pre-matches energy supply and demand before power fluctuations occur, significantly reducing the hysteresis effect of traditional feedback control. Simultaneously, the dynamic feedback correction mechanism, through multi-dimensional deviation analysis and adaptive parameter adjustment, forms a closed-loop optimization of the prediction model and control strategy, enabling the system to adapt to complex environmental changes and ensuring that the equipment operating parameters remain within a safe and efficient range.

[0017] The technical solution of this invention breaks through the excessive reliance on real-time feedback in traditional methods by using a spatiotemporal data-driven prediction-control-correction collaborative architecture. It constructs a power regulation system that combines foresight and adaptability, effectively improving the operational stability of photovoltaic hydrogen production systems under shading interference, reducing the impact of power surges on equipment lifespan, and achieving synergistic optimization of renewable energy consumption efficiency and hydrogen production system reliability.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the following describes the preferred embodiments of the present invention in detail with reference to the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is a flowchart of a photovoltaic hydrogen production power feedforward control method in one embodiment of the present invention; Figure 2 This is a schematic diagram of a spatiotemporal data acquisition system in one embodiment of the present invention; Figure 3 This is a schematic diagram of a shadow spatiotemporal prediction model in one embodiment of the present invention; Figure 4 This is a schematic diagram of a power feedforward control strategy in one embodiment of the present invention; Figure 5 This is a schematic diagram of a dynamic feedback correction mechanism in one embodiment of the present invention; Figure 6 This is a schematic diagram of the spatial correlation of weight allocation in one embodiment of the present invention; Figure 7 This is a schematic diagram of a performance comparison radar in one embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] The present invention provides the following preferred embodiments: Example 1: To address the coordination issue between power fluctuations and equipment operational stability caused by dynamic changes in shading in photovoltaic hydrogen production systems, this example provides a photovoltaic hydrogen production power feedforward control method based on shading spatiotemporal prediction, such as... Figure 1 As shown, the control method flow is as follows: S100: Construct a spatiotemporal data acquisition system, acquire solar irradiance, atmospheric temperature, cloud movement trajectory and component surface shadow distribution data of the photovoltaic array area through a multi-dimensional sensor network, divide the photovoltaic array into multiple monitoring units and perform spatiotemporal alignment processing on the real-time data to form a standardized dataset containing spatial location, timestamp and multi-dimensional physical quantities.

[0022] S200: Establish a shadow spatiotemporal prediction model, process the standardized dataset based on the prediction architecture that integrates spatiotemporal features, extract the spatial neighborhood shadow features and time series migration patterns of the monitoring unit, and generate the shadow coverage probability and intensity prediction values ​​of each monitoring unit within a set time period in the future.

[0023] S300: Formulate a power feedforward control strategy, calculate the power fluctuation range of the photovoltaic array based on the predicted value, generate feedforward control commands in combination with the power response characteristics of the hydrogen production equipment, and control the fluctuation range within the safe operating range of the equipment by adjusting the input power parameters of the electrolyzer.

[0024] S400: Constructs a dynamic feedback correction mechanism to collect the actual power generation of the photovoltaic array and the operating parameters of the hydrogen production equipment in real time, calculate the deviation between the actual power generation of the photovoltaic array and the predicted and control target parameters, and dynamically correct the predicted model parameters and control strategy compensation parameters when the deviation exceeds the preset threshold, forming a closed-loop optimization control system.

[0025] Specifically, the first step is to construct a spatiotemporal data acquisition system, such as... Figure 2 As shown, by deploying a multi-dimensional sensor network including solar irradiance sensors, atmospheric temperature sensors, cloud trajectory monitoring radar, and high-resolution shadow distribution cameras in the photovoltaic array area, the photovoltaic array is divided into spatially correlated monitoring units according to geographical coordinates, such as... Figure 2The location of the monitoring unit is marked with a hexagon. The multidimensional data collected in real time by each sensor is spatiotemporally aligned based on a timestamp synchronization algorithm and a spatial coordinate mapping model. This associates each data point with a unique three-dimensional spatial coordinate (X,Y,Z), a nanosecond-level timestamp, and physical quantities such as solar irradiance, temperature, and shadow grayscale value, forming a standardized spatiotemporal dataset that provides a unified input format for subsequent processing.

[0026] Furthermore, a spatiotemporal prediction model for shadows is established, such as... Figure 3 The illustrated spatiotemporal feature prediction architecture employs a cascaded structure of a spatial convolutional neural network (CNN) and a long short-term memory network (LSTM). Specifically, the spatial convolutional layer extracts the spatial correlation (e.g., ) between the monitoring unit and its neighborhood shadow distribution using a two-dimensional convolutional kernel. Figure 3 The spatial feature heatmap (showing neighborhood weight allocation) captures the diffusion and shading intensity transmission patterns of shadows on the photovoltaic array surface; the time-series LSTM layer performs temporal modeling on historical spatiotemporal data, identifying periodic fluctuations and anomalous mutation patterns in shadow changes. Through joint learning of spatiotemporal features, the system outputs the shadow coverage probability and intensity prediction values ​​for each monitoring unit within a set future time step, providing forward-looking data support in the time-space dimension for power fluctuation analysis.

[0027] Furthermore, in the power feedforward control strategy formulation stage, such as Figure 4 As shown, the spatiotemporal distribution of power generation of each monitoring unit in the photovoltaic array is calculated based on the shading prediction results, and the overall power fluctuation range is extrapolated through the superposition effect. Figure 4 The shaded area represents the normalized power fluctuation range. Combining the power response characteristic curve of the hydrogen production equipment with safe operating voltage and current thresholds, a feedforward control command containing power adjustment amplitude and time gradient is generated to drive the dynamic adjustment of the electrolyzer's input power parameters. This adjustment process is achieved through a piecewise linear adjustment algorithm, ensuring that the equipment's power response curve and the photovoltaic power fluctuation curve are phase-matched on the time axis, controlling the fluctuation amplitude within the equipment's safe operating boundaries. Figure 4 The restricted range shown in the power regulation curve of the electrolytic cell.

[0028] Furthermore, when constructing a dynamic feedback correction mechanism, such as Figure 5The three-dimensional deviation analysis surface shown collects real-time data on the actual power generation of the photovoltaic array and the operating parameters of the hydrogen production equipment, such as voltage, current, and temperature. A multi-dimensional deviation calculation model evaluates the spatial location deviation, time series deviation, and physical quantity deviation between the predicted and actual values. When any dimension deviation exceeds a preset threshold, the model parameter adaptive adjustment module and the control strategy compensation module are triggered to dynamically correct the adjustment coefficients of the convolution kernel weights, LSTM forget gate parameters, and control commands of the prediction model. This closed-loop mechanism achieves real-time calibration of the predictive control process through a deviation feedback path, forming a closed-loop optimization system of data acquisition, predictive control, and deviation correction, thus improving the system's long-term adaptability to complex shadow environments.

[0029] Through this embodiment, each technical component forms an organic whole based on the spatiotemporal data structure, prediction architecture, control strategy, and correction mechanism shown in the attached figures. A closed-loop control link from environmental perception to control execution is established in the photovoltaic hydrogen production system, providing an engineerable technical solution for solving the power stability control problem under the influence of shadows.

[0030] Example 2: To address the issues of accuracy and data reliability in shading monitoring of photovoltaic array areas, this example further optimizes the deployment architecture and data acquisition mechanism of the multi-dimensional sensor network. Specifically, considering the spatial layout characteristics of the photovoltaic array, a multi-modal sensing system is constructed using a spectroradiometer, an infrared thermal imager, and a dynamic vision sensor. The spectroradiometer is used to collect the radiation intensity distribution at different wavelengths in real time to quantify the impact of shading on the spectral response of the photovoltaic modules; the infrared thermal imager obtains the surface temperature field of the modules through non-contact temperature measurement to identify local hot spot anomalies caused by shading; and the dynamic vision sensor captures dynamic environmental images at a high frame rate to track the movement trajectory and shape changes of the shadows.

[0031] like Figure 2 The schematic diagram of the spatiotemporal data acquisition system shows that the monitoring units are divided according to the principle of spatial continuity. Based on the physical layout of the photovoltaic array, it is divided into several geographically adjacent regions with continuous spatial coordinates, such as the monitoring units marked with coordinates (2.5, 2), (7.5, 4), and (5, 6) in the diagram. Each monitoring unit is equipped with at least three sensors as cross-validation data acquisition points. These acquisition points are spatially triangularly distributed to ensure synchronous acquisition of multi-dimensional data from the same region. It is important to understand that the cross-validation mechanism, through mutual calibration of measurement data from different types of sensors, can effectively reduce the measurement error of a single sensor. For example, when a spectroradiometer detects an abnormal radiation intensity, the temperature data from an adjacent infrared thermal imager can be correlated with the image data from a dynamic visual sensor for verification, avoiding misjudgments caused by sensor malfunction or environmental noise.

[0032] Furthermore, the construction of the standardized dataset, in addition to including traditional real-time monitoring data such as irradiance and temperature, also specifically incorporates auxiliary information such as photovoltaic module type, installation tilt angle, and historical shading event records. The photovoltaic module type determines its spectral response characteristics and shading sensitivity. For example, polycrystalline silicon and monocrystalline silicon modules exhibit different power attenuation characteristics under the same shading conditions; the introduction of this parameter can improve the targeting of the prediction model. The installation tilt angle affects the geometric relationship of shading projection. By recording the tilt angle parameters of the modules within each monitoring unit, the spatial projection range of the shading can be calculated based on the solar altitude angle, providing a geometric correction basis for subsequent spatiotemporal prediction models. Historical shading event records contain the time of occurrence, duration, impact range, and corresponding power fluctuation data of the events, forming a time-series knowledge base used to train the long-term dependency features of the shading prediction model.

[0033] Furthermore, in the specific implementation of sensor deployment, such as Figure 2 The 3D spatial distribution shown in the code, with sensor position coordinates such as (1, 1, 1.2) and (3, 3, 0.8), is optimized to ensure vertical coverage of the photovoltaic module surface and a certain space above it, capturing shadow-forming elements at different heights, such as bird flight and vegetation swaying. The spatial mapping relationship between the monitoring unit and the sensor is achieved through a grid partitioning algorithm. Sensor nodes within each unit synchronize data in real time via a wireless sensor network, forming a spatiotemporally aligned dataset. It should be noted that the data acquisition frequency is matched to the sensor characteristics: the dynamic vision sensor acquires images at a rate of 30 frames per second, the infrared thermal imager generates a temperature matrix every 2 seconds, and the spectroradiometer outputs spectral data in real time at 50ms intervals. Synchronous fusion of multi-source data is achieved through timestamp calibration.

[0034] The advantage of this embodiment lies in the construction of a real-time data acquisition system with cross-validation capabilities through the collaborative deployment of multi-dimensional sensors and the division of spatial continuity monitoring units. Simultaneously, prior knowledge such as component characteristics, installation parameters, and historical events is integrated into a standardized dataset, providing multi-dimensional input features for subsequent shading spatiotemporal prediction models. This architectural design not only improves the reliability of field data but also enhances the semantic value of the data through the introduction of auxiliary information, enabling the system to more accurately identify shading characteristics in different scenarios and providing a more reliable decision-making basis for the power feedforward control of photovoltaic hydrogen production systems. Figure 2 The example of sensor spatial distribution and monitoring unit division shown can be used to adjust sensor density and unit boundaries according to the actual size of the photovoltaic array and terrain conditions during implementation, so as to ensure the integrity of monitoring coverage and the effectiveness of data acquisition.

[0035] Example 3: To address the accuracy issue of spatiotemporal feature coupling analysis in shadow prediction, this example further refines the prediction architecture design that integrates spatiotemporal features. Considering the coupling characteristics of spatial distribution and temporal evolution during the shadow migration process of photovoltaic arrays, the prediction architecture constructs a hierarchical feature processing system, such as... Figure 3 The schematic diagram of the shadow spatiotemporal prediction model shown is composed of a spatial feature extraction unit and a time series analysis unit. The two units achieve multidimensional representation of the shadow dynamic process through a feature fusion mechanism.

[0036] At the spatial feature processing level, the spatial feature extraction unit uses a hierarchical convolution method to parse the neighborhood shadow features of the monitoring unit. Specifically, the input layer receives data from... Figure 2 The standardized dataset of the spatiotemporal data acquisition system shown includes the spectral radiation distribution, temperature field matrix, and dynamic visual image sequence of each monitoring unit. The first convolutional layer uses convolutional kernels of different scales, such as 3×3 and 5×5 kernels, to perform parallel feature extraction on the spatial data of the monitoring unit and its adjacent areas, capturing shadow morphology features of different ranges, such as edge gradients and the contours of temperature anomaly areas. The pooling layer performs spatial dimensionality reduction on the convolutional output, preserving key features while reducing computational complexity. It is important to understand that the hierarchical convolutional structure allows the model to aggregate spatial correlation information from local to global layers. For example, by processing the first layer's output through a second convolutional layer, it is possible to identify shadow diffusion patterns across monitoring units, such as... Figure 3 The highly correlated regions shown in the spatial feature heatmap correspond to key locations in the photovoltaic array that are susceptible to the influence of adjacent shadows.

[0037] The time series analysis unit is built on a gated recurrent mechanism, and its core components employ modified gated recurrent units (GRUs) or long short-term memory networks (LSTMs), such as... Figure 3 The architecture design of the LSTM layer (time series) is as follows. This unit receives the temporalized feature vector output by the spatial feature extraction unit and selectively memorizes the historical state of shadow migration through a gating mechanism, such as the radiation intensity decay trend and shadow movement speed of a monitoring unit over a continuous period. The gating loop mechanism can effectively capture long-distance dependencies in the time series, such as seasonal shadows like periodic occlusion caused by tree ring changes or sudden shadows like the temporal feature differences of flocks of birds flying by. Furthermore, while generating shadow prediction results, the model output layer simultaneously outputs a prediction confidence index. This index quantifies the probability distribution of multi-class prediction results through a Softmax function, providing a reliability reference for subsequent control strategies—when the confidence level is lower than a preset threshold, the system automatically triggers... Figure 5 The dynamic feedback correction mechanism shown uses real-time sensor data to correct the prediction results.

[0038] In practical implementation, the input data of the prediction architecture needs to undergo spatiotemporal alignment preprocessing to... Figure 2 Asynchronous data acquisition from various sensors is converted into a unified spatiotemporal sequence matrix through timestamp calibration and spatial coordinate mapping. During training, a dataset containing historical shadow event records is used. End-to-end supervised learning optimizes convolutional layer weights and recurrent unit parameters. The loss function combines mean squared error (MSE) and confidence cross-entropy to ensure a balance between prediction accuracy and uncertainty assessment. Notably, a feature concatenation layer is placed between the spatial feature extraction unit and the time series analysis unit to fuse spatial convolutional features with the time series latent states, forming a composite feature vector containing spatiotemporal coupling information. This provides the foundation for multimodal prediction in the output layer.

[0039] The advantage of this embodiment lies in the fact that, through the collaborative design of hierarchical convolution and gated recurrent mechanisms, a predictive architecture capable of simultaneously analyzing the spatial distribution characteristics and temporal evolution patterns of shadows is constructed, and its output confidence index provides a reliability measure for system decision-making. This architectural design not only improves the spatiotemporal resolution of shadow prediction but also enhances the model's adaptability to complex occlusion scenarios through feature fusion mechanisms, such as... Figure 3 The multi-dimensional error analysis shown in the time series prediction error distribution provides a quantitative basis for subsequent parameter adjustment of the power control strategy. Through comparison with... Figure 2 Through the collaborative work of the spatiotemporal data acquisition system, this predictive architecture can generate shadow prediction results with confidence assessment in real time, providing accurate input signals for the feedforward control of the photovoltaic hydrogen production system.

[0040] Example 4: To address the coordinated control problem between power fluctuations and equipment operational stability in photovoltaic hydrogen production systems, this example further optimizes the architecture design of the power feedforward control strategy. To address the uncertainty in photovoltaic output power caused by shading, the strategy constructs a closed-loop control system including a power compensation threshold determination module, such as... Figure 4 The schematic diagram of the power feedforward control strategy shown in the figure is based on generating reasonable power adjustment commands through multi-parameter coupling analysis, while satisfying energy conservation constraints and equipment operating boundary conditions.

[0041] The core logic of the power compensation threshold determination module is based on Figure 3The module predicts the shading probability, shading intensity (such as the radiation attenuation ratio), and current electrolytic cell load rate of the output architecture to generate a power adjustment coefficient. In practice, a three-dimensional parameter mapping relationship is first established using fuzzy logic or a rule engine: when the shading probability is higher than 60% and the radiation attenuation exceeds 40%, if the equipment load rate is in the mid-to-high range (e.g., exceeding 70% of the rated power), the module triggers a first-level compensation command. The adjustment coefficient is calculated based on the energy conservation equation using a preset formula to ensure a dynamic balance between photovoltaic output fluctuations and electrolytic cell input power. It's important to understand that this module introduces a hard constraint: the equipment's rated power boundary condition. For example, the electrolytic cell input power must not be lower than 30% or exceed 90% of the rated power. Figure 4 The dashed line indicating the "equipment boundary" represents the safe operating range, preventing equipment damage caused by extreme adjustments.

[0042] The adjustment of the input power of the electrolytic cell follows a segmented adjustment strategy, such as... Figure 4 As shown in the curve of the "Electrolyzer Power Segmented Adjustment Strategy," the power change process is divided into a rapid response segment, a smooth transition segment, and a steady-state maintenance segment. In the rapid response segment, such as the initial 0-10 seconds of a sudden shading event, the adjustment rate is dynamically adjusted according to the rate of change of shading intensity, and a proportional-integral (PI) controller is used to quickly converge the power deviation. Upon entering the smooth transition segment, a slope limiter is introduced to control the power change rate and prevent current surges from damaging the electrolyzer membrane electrodes. Once the power reaches a new steady-state value, the maintenance segment uses real-time monitoring... Figure 2 The sensor data ensures that the adjusted power value continuously meets the energy conservation condition—that is, the difference between the photovoltaic output power and the electrolyzer input power does not exceed the buffer capacity of the energy storage device.

[0043] Furthermore, the segmented adjustment strategy divides the adjustment range in real time based on the equipment load rate: when the load rate is below 50%, a relatively large power fluctuation is allowed, and the upper limit of the adjustment rate is set at 5% / second of the rated power to fully utilize the photovoltaic redundant energy; when the load rate is above 80%, the adjustment rate is strictly limited to below 2% / second, and the energy storage system is triggered to work in coordination to ensure that the power change is within the safe range of the equipment. This hierarchical control mechanism combines... Figure 4 The system predicts power fluctuation range and equipment operating characteristics. For example, when the predicted power falls into the red boundary area, the system automatically switches to conservative adjustment mode to ensure equipment operation safety by reducing the adjustment amplitude coefficient.

[0044] At the hardware implementation level, power regulation commands are transmitted to the electrolytic cell controller via industrial Ethernet. The controller then adjusts the input voltage amplitude using pulse width modulation (PWM) technology to achieve smooth power regulation. It should be noted that a real-time feedback verification process is embedded in the strategy to analyze the deviation between the actual adjusted power value and the predicted value. Figure 5The dynamic feedback correction mechanism shown triggers a local prediction model update when the deviation exceeds 5%, ensuring the real-time adaptability of the control strategy. The system also features a manual / automatic switching mode, allowing maintenance personnel to intervene and adjust the amplitude coefficient in extreme shadow scenarios, enhancing the flexibility of the control strategy.

[0045] The advantage of this embodiment lies in constructing a feedforward control strategy that balances power fluctuation suppression and safe equipment operation through a multi-parameter coupled threshold determination and segmented adjustment mechanism. This strategy not only considers the quantitative indicators of shadow prediction results but also deeply integrates equipment physical characteristics and energy conservation constraints, such as... Figure 4 The power smoothing effect shown by the feedforward control curve effectively reduces the system impact caused by abrupt adjustments in traditional control methods. By correlating the adjustment amplitude coefficient with load rate and shading characteristics, the control strategy achieves scenario adaptation, providing an engineering solution for the stable operation of photovoltaic hydrogen production systems in complex shading environments.

[0046] Example 5: To address the issue of dynamic adaptation between the prediction model and the actual operating scenario, this example further refines the specific implementation of the dynamic feedback correction mechanism. For example... Figure 5 As shown, the deviation calculation module of the dynamic feedback correction mechanism integrates a multi-dimensional analysis architecture for power deviation and equipment status parameter deviation. Power deviation is represented by the difference between the real-time collected output power of the photovoltaic hydrogen production system and the predicted power. Equipment status parameter deviation covers the degree of deviation between key operating parameters of the electrolyzer, such as electrolyte temperature, pressure, and conductivity, and preset safety thresholds. It is important to understand that the aforementioned multi-dimensional deviation data is not simply superimposed, but rather transformed into a unified system health index by constructing a nonlinear mapping model. This index serves as the core basis for correction decisions and comprehensively reflects the degree of fit between the equipment operating status and the predictive control objectives.

[0047] Furthermore, the dynamic correction process is based on the collaborative processing of equipment operating status feedback data and historical correction records. Equipment operating status feedback data is acquired in real time by high-precision sensors deployed in the electrolyzer. Electrolyte temperature is measured using a platinum resistance sensor, pressure parameters are acquired using a piezoresistive transmitter, and conductivity is monitored online using an electromagnetic induction sensor. This data is transmitted to the control unit in real time via industrial Ethernet, forming a real-time feedback loop. The historical correction record module stores details of correction events within a preset time period, including the timestamp of each parameter adjustment, the parameter set before adjustment, the correction magnitude, and the operational effect feedback after adjustment, providing historical experience for current correction decisions. Based on this, the preset optimization algorithm employs a constrained gradient descent strategy. By setting a correction frequency limit factor, the maximum number of corrections per unit time is controlled within a reasonable range, avoiding the risk of system oscillation due to frequent adjustments. It is important to note that the optimization algorithm does not globally update all model parameters when adjusting parameters. Instead, it differentiates the parameter correction weights through the device health model. For devices with health indicators below the threshold, the correction weights of their corresponding parameters are increased, and control parameters that have a significant impact on the device's operating status are adjusted first, forming an adaptive correction strategy based on the device's health status.

[0048] In its implementation, the dynamic feedback correction mechanism visualizes the mapping relationship between multi-dimensional deviations and system health through a three-dimensional surface model, such as... Figure 5 As shown, the surface consists of independent variables such as power deviation and temperature deviation, and dependent variables such as health index, which can intuitively present the degree of influence of different deviation combinations on the system's operating status. The selection of correction points is based on surface gradient analysis, prioritizing deviation combinations in high curvature regions as trigger conditions for dynamic correction, ensuring that correction actions target the weak points in system operation. Furthermore, the storage and retrieval of historical correction records utilizes time-series database technology, supporting rapid querying and pattern recognition based on time windows, providing efficient data support for the optimization algorithm.

[0049] The advantage of this embodiment is that by constructing a multi-dimensional deviation analysis framework and a correction strategy driven by equipment health, dynamic optimization of the prediction model and control strategy parameters is achieved. This ensures the system's responsiveness to real-time operational deviations and avoids excessive adjustment of control parameters through correction frequency constraints, thus forming a closed-loop control system that balances adaptability and stability.

[0050] Example 6: Figure 7 As shown, to improve the response speed and prediction accuracy of the forecasting model to regional weather changes, this embodiment further optimizes the data input architecture of the spatiotemporal data acquisition system. For example... Figure 2As shown, the spatiotemporal data acquisition system, based on the existing local sensor network, integrates meteorological satellite remote sensing data and ground meteorological station observation data to construct a multi-scale meteorological information fusion mechanism. The meteorological satellite data includes macroscopic meteorological parameters such as cloud imagery information in the visible and infrared bands and solar radiation intensity inversion values, while the ground meteorological stations provide localized, refined data such as real-time wind speed, wind direction, atmospheric temperature, and humidity. The two types of data complement each other through a data fusion algorithm, overcoming the limitations of a single data source in terms of coverage and spatiotemporal resolution.

[0051] Furthermore, the data fusion algorithm employs a state estimation framework as its core processing unit. This framework establishes state-space models of external variables such as cloud movement speed and solar incidence angle to achieve collaborative processing of macroscopic meteorological information and local sensor data. Specifically, for cloud movement speed, the algorithm utilizes sequential image analysis of satellite cloud images, combines optical flow to calculate the cloud's motion vector, and corrects it using wind speed data from ground meteorological stations to form a regional-scale cloud movement prediction model. For solar incidence angle, the algorithm integrates solar position parameters provided by satellites with tilt irradiance data measured by local sensors, and estimates the real-time incidence angle using a Kalman filter algorithm to reduce measurement errors caused by atmospheric refraction, obstruction, and other factors. It is important to understand that the state estimation framework adopts a hierarchical processing structure. First, it performs time synchronization and spatial registration of different data sources, transforming the latitude and longitude coordinates of satellite data with the local coordinate system of ground sensors to ensure that the data is fused under a unified spatiotemporal reference.

[0052] Furthermore, in the specific implementation of data fusion, the system deploys a dedicated data interface module to receive broadcast data from meteorological satellites and network transmission data from ground meteorological stations in real time. After format parsing and quality verification, the data is stored in a distributed database. The preprocessed meteorological data, along with data on light intensity, temperature, humidity, etc., collected by local sensors, are then integrated into the data. Figure 2 Information acquired by sensor nodes distributed in three dimensions is collectively input into the prediction model, forming a composite input vector that includes macro-weather trends and local micro-environmental characteristics. This fusion architecture enables the prediction model to capture the movement trajectory of regional weather systems, such as shadow changes caused by cumulus cloud passage, and microclimate differences caused by local terrain and equipment layout, thereby improving the prediction accuracy of photovoltaic array shading.

[0053] Furthermore, an anomaly detection mechanism is implemented during the data fusion process. When the deviation between satellite data and ground observation data exceeds a preset threshold, a data verification process is automatically triggered, ensuring the integrity of the input data through methods such as historical data interpolation or data substitution from neighboring stations. The parameters of the state estimation framework are updated using an online learning mechanism, dynamically adjusting model weights based on the real-time fusion results to ensure adaptability to different seasons and climate regions.

[0054] Through this embodiment, the spatiotemporal data acquisition system achieves an organic combination of macro-meteorological trends and local environmental characteristics. The data fusion algorithm not only improves the representation accuracy of external variables, but also establishes physical correlations between multi-source data through the state estimation framework, providing richer and more reliable input information for the shadow spatiotemporal prediction model, thereby enhancing the adaptability of the entire photovoltaic hydrogen production system to complex meteorological conditions.

[0055] Example 7: To address the issue of insufficient real-time calculation efficiency of control parameters in photovoltaic hydrogen production systems under complex operating conditions, this example further refines the specific implementation method of the pre-planning stage. This stage involves constructing a typical scenario library containing multi-dimensional features, classifying and clustering historical operating data according to weather type, shadow duration, and coverage area to form a standardized scenario set. It is important to understand that weather type encompasses different lighting conditions such as cloudy, rainy, and partial shading; shadow duration is divided into instantaneous shading, short-term shading, and long-term shading; and coverage area is quantified based on the spatial proportion of the photovoltaic array's shaded area. These three factors together constitute the core classification dimensions of the scenario library.

[0056] During the scenario library construction process, for each scenario type, the system pre-generates corresponding control parameter combinations through offline simulation and historical operating data fitting, including key parameters such as the electrolyzer power adjustment range, voltage and current thresholds, and feedforward control coefficients. The generation of these parameter combinations is based on the coupled analysis of the photovoltaic array output characteristics and the electrolyzer dynamic response characteristics, ensuring a balance between system stability and energy efficiency under specific scenarios. Furthermore, during real-time operation, such as... Figure 3 The output of the shadow spatiotemporal prediction model shown will be matched with scene features in the scene library. The typical scene with the highest matching degree will be quickly located through the pattern recognition algorithm, and the pre-planned combination of control parameters will be directly called to avoid the delay problem caused by real-time calculation.

[0057] It's important to note that the scenario library is not statically stored but continuously updated through an online learning mechanism. When a new scenario not yet included in the library emerges, the system automatically collects real-time data and extracts its features. After verification by an expert system, the scenario is incorporated into the library, forming a closed-loop optimization system. This mechanism, combining pre-planning with real-time matching, reduces the complexity of online computation and enhances the adaptability of the control strategy through in-depth utilization of historical data.

[0058] The advantage of this embodiment is that by constructing a structured library of typical scenarios, complex and ever-changing actual working conditions are transformed into pre-planned standard scenarios, enabling rapid access to control parameters and efficient connection between system response, thus providing a systematic solution for the stable operation of photovoltaic hydrogen production systems in dynamic shading environments.

[0059] Example 8: To address the balance between equipment operational safety and control accuracy during power compensation, this example details the logical architecture of the power compensation threshold determination module. The integrated logic judgment unit uses shadow prediction results and equipment operating status parameters as input variables. The shadow prediction results include spatiotemporal characteristics such as shadow movement trajectory and coverage change rate, while the equipment operating status parameters cover key indicators such as real-time voltage, current, temperature, and pressure of the electrolytic cell. The output variables are adjustment commands for the electrolytic cell's operating voltage and current. The command generation process fully considers the equipment's thermal response characteristics and power surge constraints to avoid equipment damage or operational anomalies caused by sudden parameter changes.

[0060] Furthermore, in the logical judgment process, a thermal response model for the equipment is first established. This model, based on the electrochemical reaction kinetics within the electrolytic cell, quantifies the dynamic relationship between voltage and current adjustments and equipment temperature rise, ensuring that the adjustment range remains within the equipment's thermal tolerance range. For power surge constraints, the system presets an upper limit for the rate of change of current and a voltage fluctuation threshold. When the predicted power fluctuation caused by shading exceeds the threshold, the logic unit generates a step-by-step adjustment command through a segmented adjustment strategy, rather than directly making a large-scale adjustment, thereby avoiding the instability of the electrochemical process caused by power surges. Furthermore, the judgment unit uses a fuzzy logic algorithm to handle the coupling relationship of multi-variable inputs, transforming factors such as shading intensity, current equipment load status, and historical adjustment effects into fuzzy rules, forming a non-linear command generation logic, improving the rationality of decision-making under complex operating conditions.

[0061] It is important to understand that the real-time acquisition of equipment operating status parameters depends on... Figure 2 The spatiotemporal data acquisition system shown achieves high-precision synchronous monitoring of multiple physical quantities through a sensor network distributed across the photovoltaic array and hydrogen production equipment, providing a reliable data foundation for logical judgments. The output of the adjustment command is then... Figure 4 The power feedforward control strategy shown forms a closed loop. The feedforward control plans the power adjustment direction in advance based on the prediction results, while the dynamic adjustment command of this module serves as feedback correction. The combination of the two achieves precise control of the equipment operating parameters.

[0062] The advantage of this embodiment is that by integrating a logic judgment unit that takes into account the physical characteristics of the equipment, the shadow prediction information is deeply integrated with the actual operating state of the equipment, and the power compensation is finely adjusted under the premise of ensuring the safe operation boundary of the equipment, thus providing a mechanism guarantee for the reliable operation of the photovoltaic hydrogen production system.

[0063] Example 9: To address the issue of insufficient correlation analysis between adjacent region features during shadow prediction, this example elaborates on the weight allocation mechanism for shadow coverage probability and intensity prediction. This mechanism, based on the spatial location correlation of monitoring units, constructs a composite weight matrix including distance weight and historical co-occurrence frequency weight to strengthen the analysis of mutual influence between shadow features in adjacent regions, thereby improving the prediction model's ability to capture spatial distribution patterns.

[0064] Furthermore, based on Figure 2 The sensor network layout shown determines the spatial coordinates of each monitoring unit, calculates the Euclidean distance between any two monitoring units, and generates a distance weight matrix. Monitoring units that are closer together have higher weight coefficients due to the spatial continuity of shadow propagation, and vice versa, reflecting the degree of correlation in physical space. Furthermore, the historical co-occurrence frequency weight is constructed by analyzing the frequency of simultaneous shadow occurrences of different monitoring units in historical data. Monitoring units with high-frequency co-occurrence are assigned higher weights to reflect the regularity of shadow distribution over long-term operation. The two weights are then linearly combined after normalization to form the final spatial correlation weight, which is used to correct the prediction model for shadow coverage probability and intensity.

[0065] Furthermore, during the operation of the prediction model, such as Figure 6 The spatial correlation mechanism of weight allocation shown is embedded in the spatiotemporal data processing stage. When processing real-time data of a certain monitoring unit, the model not only considers the independent monitoring value of that unit, but also aggregates the historical data and real-time status of adjacent units through the weight matrix to form an input vector containing spatial correlation features. This processing method can effectively capture the spatial propagation trend and coordinated change pattern of shadows, avoiding the one-sidedness of prediction based on single unit data. It is important to understand that the weight matrix is ​​not fixed, but is dynamically updated according to the latest monitoring data through an online learning algorithm to adapt to changes in shadow distribution patterns caused by environmental changes.

[0066] Furthermore, combined Figure 3 The shadow spatiotemporal prediction model shown here uses a weight allocation mechanism as a key step in spatial feature extraction, which is deeply integrated with time series analysis modules such as LSTM layers to form a spatiotemporal dimension. The introduction of spatial correlation weights enables the model to more accurately identify the spatial patterns of shadow distribution when processing multidimensional spatiotemporal data, providing a more reliable prediction basis for subsequent power feedforward control and dynamic feedback correction.

[0067] The advantage of this embodiment is that by quantifying the spatial correlation between monitoring units, a weight allocation system that conforms to the physical laws of shadow propagation is constructed, which improves the adaptability of the prediction model to complex shadow environments and lays a data processing foundation for the precise control of photovoltaic hydrogen production systems.

[0068] Example 10: To address the systematic construction of multi-module collaborative control in photovoltaic hydrogen production systems, this example further refines the system architecture implementation of the corresponding control method. This system achieves functional decoupling and process integration through modular design. Each module independently executes its specific function while forming an organic whole through data interaction, ensuring closed-loop operation throughout the entire process from data acquisition to control correction.

[0069] Furthermore, the data acquisition and preprocessing module serves as the system input interface, based on... Figure 2 The sensor network layout shown employs multi-dimensional sensors deployed within the photovoltaic array area, including light intensity sensors, weather radar, and infrared monitoring devices, to acquire spatiotemporal data in real time. This module first divides the photovoltaic array into several monitoring units, each corresponding to independent spatial coordinates and monitoring range. A spatiotemporal alignment algorithm is used to unify the timestamps and spatially calibrate the asynchronous data from different sensors, generating a standardized dataset containing parameters such as time series, geographic coordinates, and light intensity. It's important to understand that the module also integrates external meteorological data, such as satellite cloud images and regional weather forecasts, using data fusion algorithms to correct the limitations of local monitoring data and improve the completeness and reliability of the input data.

[0070] Furthermore, the shadow spatiotemporal prediction module is built upon a prediction architecture that integrates spatiotemporal features, such as... Figure 3 As shown, it integrates a spatial feature extraction unit and a time series analysis unit. The spatial feature extraction unit adopts the weight allocation mechanism described in Example 9, constructing a spatial correlation matrix based on the distance between monitoring units and historical coordination frequency to quantify the mutual influence of shadow features in adjacent areas. The time series analysis unit processes historical data of a single monitoring unit through a recurrent neural network such as LSTM to capture the evolution of shadow intensity over time. After the spatiotemporal feature vectors output by the two units are processed by the fusion layer, the shadow coverage probability and intensity prediction values ​​of each monitoring unit are generated, providing a decision-making basis for the subsequent control module.

[0071] Furthermore, the power feedforward control module takes the predicted value as input and first calculates the power fluctuation range by combining the power compensation threshold determination logic described in Example 8 with the electrolytic cell equipment operating status parameters such as real-time voltage, current, and temperature. Considering the constraints of sudden power changes in the equipment, the module has a built-in segmented adjustment mechanism that breaks down power adjustment commands exceeding the threshold into multi-level stepped outputs to avoid drastic impacts on the electrolytic cell's operating status. During the command generation process, the module synchronously calls the pre-planned parameters from the typical scenario library described in Example 7 and dynamically adjusts the combination of control parameters according to the real-time matched scenario type, achieving efficient integration of prediction results and equipment control.

[0072] Furthermore, the dynamic feedback correction module, as a key component of closed-loop control, calculates deviations between real-time acquisition of actual electrolyzer operating parameters, such as output hydrogen flow rate and stack voltage fluctuations, and predicted values. When multi-dimensional deviations, including power deviation, voltage deviation, and time delay deviation, exceed preset thresholds, the module initiates a parameter correction process: first, the multi-dimensional deviation analysis unit locates the source of the deviation, such as spatial feature extraction errors in the prediction model or unreasonable segmented parameters of the control strategy; then, the parameter selective adjustment unit locally updates the weight matrix of the corresponding module's prediction model or the threshold parameters of the control strategy, avoiding system instability caused by global adjustments. This targeted correction mechanism ensures that the system continuously adapts to environmental changes and equipment characteristic drift during long-term operation.

[0073] The advantage of this embodiment lies in the fact that by clearly defining the architectural design and interaction logic of each functional module, a complete closed-loop system from data acquisition, predictive analysis, feedforward control to feedback correction is constructed. The specialized design of each module, such as the spatiotemporal alignment mechanism for data preprocessing, the spatiotemporal feature fusion architecture of the prediction module, the segmented adjustment strategy of the control module, and the directional parameter adjustment of the correction module, not only ensures the efficient implementation of system functions, but also provides standardized interfaces for subsequent modular upgrades and personalized configurations, forming a photovoltaic hydrogen production power control technology system adapted to complex shading environments.

[0074] Although the present invention has been specifically described above with reference to preferred embodiments, it should be understood that the present invention is not limited to the embodiments described above. Various modifications and variations can be made by those skilled in the art without departing from the spirit of the present invention, and such modifications and variations should fall within the scope defined by the appended claims and their equivalents.

Claims

1. A photovoltaic hydrogen production power feedforward control method based on shadow spatiotemporal prediction, characterized in that, The steps of the method comprise: Constructing a spatio-temporal data acquisition system, acquiring solar irradiance, atmospheric temperature, cloud movement trajectory and component surface shadow distribution data of the photovoltaic array area through a multi-dimensional sensor network, dividing the photovoltaic array into multiple monitoring units and performing spatio-temporal alignment processing on real-time data to form a standardized data set including spatial position, time stamp and multi-dimensional physical quantity; Establishing a shadow spatio-temporal prediction model, processing the standardized data set based on a prediction architecture that fuses spatio-temporal features, extracting spatial neighborhood shadow features and time series migration rules of the monitoring units, and generating shadow coverage probability and intensity prediction values of each monitoring unit within a future set time; Developing a power feedforward control strategy, calculating the photovoltaic array power generation power fluctuation range according to the prediction values, combining the hydrogen production equipment power response characteristics to generate feedforward control instructions, and adjusting the electrolytic cell input power parameters to control the fluctuation amplitude within the safe operation range of the equipment; Building a dynamic feedback correction mechanism, real-time acquisition of the actual power generation of the photovoltaic array and the operation parameters of the hydrogen production equipment, deviation calculation with the prediction and control target parameters, dynamic modification of the prediction model parameters and control strategy compensation parameters when the deviation exceeds the preset threshold, forming a closed-loop optimization control system.

2. The photovoltaic hydrogen generation power feedforward control method based on shadow spatiotemporal prediction according to claim 1, characterized in that, The multi-dimensional sensor network includes a spectral radiometer, an infrared thermal imager and a dynamic visual sensor deployed in the photovoltaic array area, the monitoring unit division is based on the principle of spatial continuity, each monitoring unit corresponds to at least three cross-validation data acquisition points of the sensors, and the standardized data set further includes photovoltaic component type, installation inclination and historical shadow event record.

3. The photovoltaic hydrogen generation power feedforward control method based on shadow spatio-temporal prediction according to claim 1, characterized in that, The prediction architecture that fuses spatio-temporal features includes a spatial feature extraction unit and a time series analysis unit, the spatial feature extraction unit processes the neighborhood shadow features of the monitoring units in a hierarchical convolution manner, the time series analysis unit captures the time correlation of shadow migration through a gated recurrent mechanism, and the output layer of the shadow spatio-temporal prediction model simultaneously outputs the confidence index of shadow prediction.

4. The photovoltaic hydrogen generation power feedforward control method based on shadow spatio-temporal prediction according to claim 1, characterized in that, The power feedforward control strategy includes a power compensation threshold determination module, the power compensation threshold determination module generates a power adjustment amplitude coefficient according to the shadow coverage probability, the shadow intensity and the equipment load rate, the electrolytic cell input power adjustment follows the energy conservation constraint and the equipment rated power boundary condition, and adopts a segmented adjustment method to control the power change rate.

5. The photovoltaic hydrogen generation power feedforward control method based on shadow spatio-temporal prediction according to claim 1, characterized in that, The deviation calculation of the dynamic feedback correction mechanism includes multi-dimensional analysis of power deviation and equipment state parameter deviation, the dynamic modification process combines equipment operation state feedback data and historical correction records, selectively adjusts the prediction model and control strategy parameters through a preset optimization algorithm, the optimization algorithm sets a correction frequency limit to avoid system oscillation; the equipment operation state feedback data includes electrolyte temperature, pressure and conductivity of the electrolytic cell, the historical correction record includes correction events and parameter adjustment records in the past preset time period, and the dynamic modification process differentiates the parameter modification weight by establishing an equipment health degree model.

6. The photovoltaic hydrogen generation power feedforward control method based on shadow spatio-temporal prediction according to claim 1, characterized in that, The spatio-temporal data acquisition system also accesses meteorological satellite and ground meteorological station data, combines macro-weather information with local sensor data through a data fusion algorithm, and uses a state estimation framework to process external variables such as cloud layer movement speed and solar incident angle, thereby improving the response capability of the prediction model to regional weather changes.

7. The photovoltaic hydrogen generation power feedforward control method based on shadow spatio-temporal prediction according to claim 1, characterized in that, The method also includes a pre-planning stage, in which a typical scenario library including different weather types, shadow duration and coverage range is established based on historical data, and a control parameter combination is generated in advance for each scenario. When the real-time prediction result matches the scenario library scenario, the corresponding pre-planned control parameter is directly called.

8. The photovoltaic hydrogen generation power feedforward control method based on shadow spatio-temporal prediction according to claim 5, characterized in that, The power compensation threshold determination module integrates a logical judgment unit, the input variables include shadow prediction results and device operating state parameters, and the output variable is the adjustment instruction of the electrolytic cell working voltage and current. The generation process of the adjustment instruction takes into account the device thermal response characteristics and power mutation constraint conditions. 9.The photovoltaic hydrogen production power feedforward control method based on shadow spatiotemporal prediction of claim 5, wherein, The generation process of the shadow coverage probability and intensity prediction value includes weight distribution of the spatial position correlation of the monitoring unit. The weight distribution is based on the distance between monitoring units and the historical frequency of cooperative appearance of shadows, in order to strengthen the mutual influence analysis of shadow characteristics in adjacent areas.

10. A shadow spatiotemporal prediction based photovoltaic hydrogen production power feedforward control system for implementing the shadow spatiotemporal prediction based photovoltaic hydrogen production power feedforward control method according to any one of claims 1-9, characterized in that, The system comprises: a data acquisition and preprocessing module that acquires spatio-temporal data of a photovoltaic array area through a multi-dimensional sensor network, divides monitoring units, and performs spatio-temporal alignment processing to generate a standardized data set; a shadow spatio-temporal prediction module that processes the standardized data set based on a prediction framework that fuses spatio-temporal characteristics, to generate shadow coverage probability and intensity prediction values for each monitoring unit; a power feedforward control module that calculates a power fluctuation range and generates a feedforward control instruction based on the prediction values, to adjust the input power parameters of the electrolytic cell to control the fluctuation amplitude; a dynamic feedback correction module that acquires actual operating parameters and performs deviation calculation, and dynamically corrects the prediction model and control strategy parameters when the deviation exceeds a preset threshold, to form a closed-loop optimization control system; The data acquisition and preprocessing module accesses external meteorological data and performs fusion processing. The shadow spatio-temporal prediction module includes a spatial feature extraction unit and a time series analysis unit. The power feedforward control module sets a power compensation threshold determination and segmented adjustment mechanism. The dynamic feedback correction module integrates a multi-dimensional deviation analysis unit and a parameter selective adjustment unit.

Images