Liftable BIPV photovoltaic roof power system based on intelligent control

Abstract

The invention relates to the technical field of photovoltaic power generation systems, and discloses a liftable BIPV photovoltaic roof power system based on intelligent control, which comprises a panoramic light field sensing module, a multi-modal data fusion prediction module, an ambient light contribution analysis module, a self-adaptive angle optimization module and a partition cooperative control module. According to the invention, through omnibearing light field perception, multi-modal data fusion prediction, ambient light contribution analysis, adaptive angle optimization and partition cooperative control, intelligent optimization control of the angle of the photovoltaic roof assembly is realized, the light energy capture efficiency can be comprehensively improved, and the system has a prospective control capability and is suitable for popularization and application. The problems that in the prior art, response to rapid change illumination conditions is lagged in consideration of direct light of a single sun track, and illumination differences of different areas in a large system exist are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation systems, and more specifically, to a liftable BIPV photovoltaic rooftop power system based on intelligent control. Background Art

[0002] With the rapid development of clean energy, photovoltaic power generation has become an important way to utilize renewable energy. In building integrated photovoltaic (BIPV) systems, the angle of photovoltaic modules has a significant impact on power generation efficiency.

[0003] However, existing photovoltaic rooftop angle control systems have three key problems: they only consider direct light from a single solar track and ignore the contributions of ambient reflected and scattered light, resulting in incomplete energy capture; they react slowly to rapidly changing lighting conditions and are unable to adjust the angle of the components in real time, resulting in inefficient light energy utilization; and they adopt a unified control strategy that is unable to cope with the lighting differences in different areas of a large system, resulting in a decrease in the overall efficiency of the system.

[0004] Existing technologies are unable to achieve forward-looking and differentiated angle control under complex lighting conditions. An intelligent control system is needed that can comprehensively consider multi-dimensional lighting environment data and make forward-looking adjustments based on the predicted results. Summary of the Invention

[0005] The present invention provides a liftable BIPV photovoltaic rooftop power system based on intelligent control, which solves the technical problems of single sun trajectory consideration, delayed response to light changes and unified control strategy in related technologies.

[0006] The present invention provides a liftable BIPV photovoltaic rooftop power system based on intelligent control, comprising: The panoramic light field perception module uses a panoramic light field perception system to collect multi-dimensional lighting environment data including direct light, scattered light, and reflected light, and generates real-time light field distribution information; The multimodal data fusion prediction module uses a multimodal data fusion prediction algorithm to process real-time light field distribution information and form a prediction model for light changes in the next 120 minutes. The ambient light contribution analysis module uses a ray tracing algorithm combined with a light change prediction model to analyze the ambient light contribution and generate a light gain model that reflects the impact of different light sources on components in each area. Adaptive angle optimization module, which executes the adaptive angle optimization algorithm to calculate the optimal angle adjustment trajectory of the photovoltaic module based on the light gain model and outputs the angle optimization instruction; The partition collaborative control module implements the partition collaborative control system, performs differentiated angle adjustments on photovoltaic modules according to angle optimization instructions, and forms an overall optimization control solution.

[0007] In a preferred embodiment, the panoramic light field perception system includes: The 360° panoramic light field sensor array is configured to collect all-round light intensity, angle and spectral distribution information; A network of ground-based light sensors deployed to collect reflected and scattered light data; Panoramic camera, satellite cloud image receiver and meteorological data collector, used to collect sky image sequences, satellite cloud images and meteorological forecast data; The data synchronization and spatial registration processing unit is used to integrate various types of data into a multi-dimensional lighting environment dataset according to time and space coordinates.

[0008] In a preferred embodiment, the multimodal data fusion prediction algorithm includes: Multi-stream convolutional long short-term memory networks, including spherical convolutional long short-term memory units, two-dimensional convolutional long short-term memory networks, deep residual convolutional long short-term memory networks, and multi-layer perceptron and long short-term memory network hybrid models, are used to process different types of input data; Adaptive attention fusion module, including feature transformation layer, weather condition encoder, time period encoder, conditional attention calculation unit, multi-head attention mechanism and feature fusion layer; The multi-scale temporal prediction network, including short-term prediction model, medium-term prediction model and long-term prediction model, is used to generate illumination prediction results for different time windows.

[0009] In a preferred embodiment, the ray tracing algorithm includes: Construct a 3D model of the environment surrounding the PV system, including spatial information and material properties of buildings, terrain, and other objects; Apply the bidirectional reflectance distribution function model to calculate the reflective characteristics of the environment surface to light; Implement the ray tracing process of multiple reflections and calculate the influence weights of different light sources on components in each area; A dynamic light gain model is constructed to quantify the contribution of environmental factors to photovoltaic power generation at different times.

[0010] In a preferred embodiment, the adaptive angle optimization algorithm includes: Establish a model for the relationship between photovoltaic module angle and power generation efficiency, describing the relationship between incident light intensity, incident angle, and module photoelectric conversion efficiency; Construct a forward-looking optimization objective function to maximize the cumulative power generation within the forecast period; A dynamic programming algorithm is applied to solve the optimal angle adjustment trajectory under the constraints of component angle adjustment speed limit, angle range limit and adjacent component spacing setting; Based on the changes in spectral characteristics, the component tilt angle is optimized to maximize energy capture in specific bands.

[0011] In a preferred embodiment, the partition collaborative control system includes: A clustering algorithm for system partitioning based on the similarity of light characteristics is used to divide the component set in the photovoltaic system into multiple areas with similar light characteristics; A partition-level angle control model to achieve coordinated adjustment of components within a region; Distributed actuator collaborative control algorithm to ensure the smoothness and accuracy of angle adjustment; Inter-regional collaborative optimization mechanism is used to ensure smooth transition of system power by controlling timing staggering.

[0012] In a preferred embodiment, the structure of the multi-stream convolutional long short-term memory network includes: Input layer, used to receive sensory data of different modalities; Multiple parallel convolutional LSTM layers are used to simultaneously capture the spatial features and temporal dependencies of different modal sensory data; Feature extraction layer, used to obtain a fixed-dimensional feature vector through global average pooling; The output layer is used to output the features extracted from each modal data.

[0013] In a preferred embodiment, the implementation of the adaptive angle optimization algorithm includes: State space construction, discretizing the angle adjustment trajectory into a grid state space with discretization levels; Adaptive time step adjustment, dynamically adjusting the time step according to the predicted rate of illumination change; The reward function is constructed by comprehensively considering multiple factors such as predicted power generation, adjusted energy consumption and mechanical stress; Forward dynamic programming, constructing a value function and recording the optimal predecessor state for each state; Reverse path extraction, backtracking from the terminal state to obtain the complete optimal angle trajectory.

[0014] In a preferred embodiment, when the multimodal data fusion prediction algorithm processes the lighting environment data, different weights are assigned to each data source according to different weather conditions and time periods: On sunny days, a higher weight is given to direct light data; In overcast or cloudy weather, higher weights are given to scattered light and satellite cloud image data; During sunrise and sunset, higher weights are given to spectral data.

[0015] In a preferred embodiment, a computer-readable storage medium is characterized in that it is used to store computer-readable instructions, which can operate a liftable BIPV photovoltaic rooftop power system based on intelligent control when the computer-readable instructions are read by a computer.

[0016] The beneficial effects of the present invention are: Comprehensively improve light energy capture efficiency and increase the total power generation of the system by comprehensively utilizing direct light, scattered light and reflected light; Achieve forward-looking control. Through multimodal data fusion prediction, the system can predict lighting changes in advance and make angle adjustments, avoiding the delayed response problem of traditional systems when lighting changes suddenly; Solve the problem of uneven regional lighting. Through zone-based differentiated control, components in different areas of a large system can independently optimize their angles based on local lighting conditions, improving overall system performance. Adapting to the changing characteristics of the spectrum, the system can dynamically adjust the component tilt according to the spectrum characteristics of different time periods, making the power generation curve more stable throughout the day; With strong adaptive capabilities, the system can automatically adjust control strategies according to seasonal and environmental changes, reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a module diagram of a liftable BIPV photovoltaic rooftop power system based on intelligent control according to the present invention; Figure 2 is a detailed flow chart of generating real-time light field distribution information of the present invention; Figure 3 It is a detailed flow chart of forming a light change prediction model for the next 120 minutes according to the present invention; Figure 4 is a detailed flow chart of the present invention for generating an illumination gain model reflecting the effects of different light sources on components in each area; Figure 5 is a detailed flow chart of the output angle optimization instruction of the present invention; Figure 6 It is a detailed flow chart of forming the overall optimization control scheme of the present invention. DETAILED DESCRIPTION

[0018] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a BIPV photovoltaic rooftop power system based on intelligent control, such as Figures 1 to 6 As shown, including: The panoramic light field perception module uses a panoramic light field perception system to collect multi-dimensional lighting environment data including direct light, scattered light, and reflected light, and generates real-time light field distribution information; It includes the following sub-steps: Step 1.1: Configure a 360° panoramic light field sensor array to collect all-round light intensity, angle, and spectral distribution information; The sensor array consists of multiple light intensity sensors, angle sensors and spectrum analyzers, which are arranged according to the principle of uniform distribution on the spherical surface to achieve light field information collection without blind spots. The measurement data of each sensor Indicates at a point in time At the incident angle Light intensity and wavelength The distribution of Indicates the The measurement data of the sensor.

[0020] Step 1.2: Deploy a ground-based light sensor network to collect reflected and scattered light data. The network consists of multiple light sensors evenly distributed around the photovoltaic system, each sensor measures the light parameters of the local area ,in, Indicates the Measurement data from ground sensors, Represents the intensity of reflected light, represents the wavelength, Indicates a time point Step 1.3: Collect sky image sequences, satellite cloud images, and weather forecast data to form auxiliary data sets; Sky images are taken at fixed time intervals by a panoramic camera Acquire and generate time series image sets ; Satellite cloud image data and weather forecast data It is updated every 10 minutes through the API interface.

[0021] Step 1.4, construct a multidimensional lighting environment dataset; Through data synchronization and spatial registration algorithms, the above-mentioned data are unified and integrated according to time and space coordinates: ; in, Indicates a time point A complete lighting environment dataset, Indicates the A panoramic light field sensor at time The collected data, Indicates the light intensity, represents the incident angle of the light, represents the wavelength, Indicates the Ground light sensors at time The collected data, Represents the intensity of reflected light, represents the wavelength, Indicates a time point The collected sky panoramic image has a resolution of Pixels, used to analyze cloud distribution and movement, Indicates a time point Obtained satellite cloud image data, Indicates a time point weather forecast data.

[0022] Data synchronization accuracy: The time synchronization error of all data sources is controlled within ±1 second to ensure the time consistency of data from different sensors; Spatial registration accuracy: The three-dimensional spatial coordinate registration error is less than 0.5 meters, ensuring the accurate correspondence between different sensor data in spatial positions; Data storage format: Adopts the hierarchical HDF5 format, supports efficient spatiotemporal indexing and querying, and facilitates the storage and rapid access of large-scale data.

[0023] The multimodal data fusion prediction module uses a multimodal data fusion prediction algorithm to process real-time light field distribution information and form a prediction model for light changes in the next 120 minutes. It includes the following sub-steps: Step 2.1: Construct a multi-stream convolutional long short-term memory network to extract the spatiotemporal features of each modal data; For each data type, a specific convolutional long short-term memory unit (ConvLSTM) is configured for processing; For light field sensor data , apply spherical convolution operation to extract direction-aware features: ; in, represents the feature vector of the light field sensor data obtained after processing by the spherical convolution long short-term memory network, represents the spherical convolutional long short-term memory network, Indicates the Measurement data from a light field sensor.

[0024] The multi-stream convolutional long short-term memory network model is one of the innovative models in this implementation. Its specific structure includes an input layer, multiple parallel convolutional LSTM layers, a feature extraction layer, and an output layer. Each modality data passes through a dedicated processing branch, which contains 3-5 layers of convolutional LSTM units, each with 16-64 convolution kernels.

[0025] Taking the sky image sequence processing branch as an example, its internal structure is: Input layer: receiving A sequence of sky images with a resolution of 12; The first convolutional LSTM layer uses 32 Convolution kernel, with a step size of 1, keeps the spatial dimension unchanged; The second convolutional LSTM layer uses 48 Convolution kernel with a stride of 2, downsampling the spatial dimension; The third convolutional LSTM layer uses 64 Convolution kernel, step size is 1; Feature extraction layer: obtains a fixed-dimensional feature vector through global average pooling; This model performs well in handling real-time changing cloud movement conditions. For example, under cloudy conditions with wind speeds of 6-8m / s, it can accurately predict cloud cover changes within the next 30 minutes.

[0026] For ground lighting data , applying a 2D convolutional long short-term memory network: ; in, Represents the feature vector of ground illumination data obtained after processing by the two-dimensional convolutional long short-term memory network, represents a two-dimensional convolutional long short-term memory network, Indicates the Measurement data from ground-based light sensors; For sky image sequences , applying deep residual convolutional long short-term memory network: ; in, represents the feature vector of the sky image sequence obtained after processing by the deep residual convolution long short-term memory network, represents a deep residual convolutional long short-term memory network, Indicates a time point Collected panoramic images of the sky; For satellite cloud images and weather data , using a hybrid model of multi-layer perceptron and long short-term memory network: ; in, Represents the satellite cloud image and meteorological data feature vectors obtained after processing by the multi-layer perceptron and long short-term memory network hybrid model. represents a hybrid model of multilayer perceptron and long short-term memory network, Indicates a time point Obtained satellite cloud image data, Indicates a time point weather forecast data.

[0027] Step 2.2: Build an adaptive attention fusion module to integrate multimodal features. This module calculates the importance weight of each feature under different weather conditions and time periods. , to achieve adaptive feature fusion: ; ; in, Indicates the The attention weight coefficient of the feature vector, 、 Respectively represent and The learnable weight matrix corresponding to the eigenvectors is Indicates the feature vectors, 、 Respectively represent and The learnable bias parameters corresponding to the feature vectors, represents the natural exponential function, For all eigenvectors Perform the sum operation, Represents the comprehensive feature vector after fusion by the attention mechanism; The adaptive attention fusion module is a key innovative module of this implementation, and its structure includes: Feature transformation layer: transforms each modal feature to the same feature space dimension; Weather Condition Encoder: Generates a condition vector based on current and historical weather conditions ; Period encoder: Divide a day into 8 periods and generate period feature vectors ; Conditional attention calculation unit: According to and Dynamically adjust the attention calculation method; Multi-head attention mechanism: 8 attention heads are calculated in parallel, each head focuses on a different feature subspace; Feature fusion layer: integrates the features of each modality according to the calculated attention weights; In practice, this module automatically adjusts the importance of each data source based on varying conditions. For example, on sunny days, direct sunlight data receives a higher weight (approximately 0.6-0.7); on overcast or cloudy days, diffuse light and satellite cloud imagery data receive a significantly higher weight (approximately 0.5-0.6); and during sunrise and sunset, spectral data receives an increased weight (approximately 0.4-0.5). This dynamic adjustment capability ensures that sunlight forecasts maintain high accuracy even in changing weather conditions.

[0028] Step 2.3: Build a multi-scale time prediction network to generate a high-precision light prediction model for the next 120 minutes. The network adopts a hierarchical prediction structure, using models of different time resolutions for short-term (0-30 minutes), medium-term (30-60 minutes), and long-term (60-120 minutes) predictions: ; ; ; in, Indicates the illumination forecast result output by the short-term forecast model, with the forecast time range being 0-30 minutes in the future. Indicates the illumination forecast result output by the medium-term forecast model, with the forecast time range being 30-60 minutes in the future. Indicates the illumination forecast result output by the long-term forecast model, with the forecast time range being 60-120 minutes in the future. represents the short-term prediction function, represents the medium-term forecast function, represents the long-term prediction function, represents the comprehensive feature vector after adaptive attention fusion, represents the time advance of the prediction, Indicates the current time point; The prediction model provides accurate input data for subsequent lighting optimization.

[0029] The ambient light contribution analysis module uses a ray tracing algorithm combined with a light change prediction model to analyze the ambient light contribution and generate a light gain model that reflects the impact of different light sources on components in each area. It includes the following sub-steps: Step 3.1, constructing a three-dimensional model of the environment surrounding the photovoltaic system; Use lidar scanning or satellite image analysis technology to obtain three-dimensional spatial information of buildings, terrain and other objects within 500 meters around the photovoltaic system and generate an accurate three-dimensional model of the environment . Material properties of each surface in the model Including reflectivity and absorption rate ,These parameters are obtained through field measurements or material databases.

[0030] Step 3.2, apply the bidirectional reflectance distribution function (BRDF) model to calculate the reflective properties of the environment surface to light; For each surface element in the environment model , according to its material properties , establish its bidirectional reflectance distribution function , describing the incident light Reflected in the outgoing direction Ratio: ; in, represents the bidirectional reflectance distribution function, represents the direction vector of the incident light, represents the direction vector of the reflected light, represents the wavelength, Indicates from point along The differential radiance of the directional radiation, Indicates from Direction incident to point The differential irradiance of point x is input as coordinates when it is used as model input.

[0031] Step 3.3: Implement a ray tracing algorithm for multiple reflections to calculate the weight of the influence of different light sources on the components in each area. The algorithm launches light from a light source (including direct sunlight and scattered light from the sky) and traces its propagation path through the environment until it reaches the photovoltaic panel or the energy decays below a threshold.

[0032] For the PV panels, from the direction The received light intensity is calculated as follows: ; in, Indicates the PV panels at a time From the direction The wavelength received is The total light intensity, Indicates time From the direction The incident wavelength is The direct light intensity, Indicates time From the surface elements After reflection, from the direction The wavelength incident on the component is The reflected light intensity, Represents all surface elements in the environment that may produce reflections Perform the summation, Indicates time From the direction The incident wavelength is The intensity of the sky light.

[0033] Step 3.4, construct a dynamic light gain model to quantify the contribution of environmental factors at different time periods; For each PV panel in the system , establish its total light gain function : ; in, Indicates the PV panels at a time The total light gain coefficient, represents the set of all possible light incident directions in the hemispherical space, Indicates the spectrum wavelength range, which is 300-1100nm, represents the total light intensity, represents the differential element of wavelength, Represents the differential element of the solid angle, integrated over the entire hemisphere Sum, Indicates the range of the incident direction of direct sunlight. It represents the main incident direction of direct sunlight, which is the solid angle corresponding to the solar disk. Indicates the intensity of direct sunlight; Indicates the The wavelength of the photovoltaic module is The light response function is obtained through laboratory testing. The specific method is: Use a monochromatic light source to scan the photovoltaic modules in the wavelength range of 300-1100nm at intervals of 10nm and record the short-circuit current at each wavelength With standard irradiance The ratio of .

[0034] For different types of photovoltaic modules (monocrystalline silicon, polycrystalline silicon, thin film, etc.), the response function curve shapes are significantly different. The system will select the corresponding response function database based on the actual installed module type; The light gain model The contribution rate of ambient light to photovoltaic power generation is quantified, providing a basis for angle optimization. Indicates that ambient light has a positive contribution, and the contribution rate is .

[0035] Adaptive angle optimization module, which executes the adaptive angle optimization algorithm to calculate the optimal angle adjustment trajectory of the photovoltaic module based on the light gain model and outputs the angle optimization instruction; It includes the following sub-steps: Step 4.1, establish a relationship model between photovoltaic module angle and power generation efficiency; For each photovoltaic module , its output power With the incident light intensity , angle of incidence and the module photoelectric conversion efficiency The relationship between can be expressed as: ; in, Indicates the PV panels at a time , azimuth is , the inclination angle is The output power when Indicates the azimuth angle of the photovoltaic module, Indicates the inclination angle of the photovoltaic module, represents the spectral wavelength range considered, represents the set of all possible light incident directions in the hemispherical space, Indicates time From the direction Incident on photovoltaic modules The wavelength is The light intensity, Indicates the incident direction of light The angle between the component normal and Represents photovoltaic modules For wavelength , the incident angle is The photoelectric conversion efficiency, Represents photovoltaic modules The effective area is obtained by measuring the physical size of the photovoltaic module and is calculated by multiplying the product of the module length and width by the fill factor (0.9-0.98, which represents the proportion of the actual photovoltaic conversion area on the module surface to the total area). represents the differential element of the solid angle, Represents the differential element of wavelength.

[0036] Step 4.2: Construct a forward-looking optimization objective function to maximize the cumulative power generation within the forecast period; For the prediction time window Inside( is 120 minutes), the angle optimization goal is: ; in, Indicates the azimuth angle and inclination The trajectory is optimized to maximize the objective function. Indicates that in the time window Inside, photovoltaic panels According to the angle trajectory and The accumulated power generation during operation, Indicates the length of the prediction time window, which is set to 120 minutes. Represents photovoltaic modules In time The azimuth adjustment trajectory, Represents photovoltaic modules In time The tilt adjustment trajectory, Represents the integral variable within the time window, ranging from , The energy consumption function of the angle adjustment is calculated by the following formula: ; in, Presentation Component The unit angle adjustment power coefficient of the azimuth drive motor is 0.5-2.0W·min / °; Presentation Component The power coefficient per unit angle adjustment of the tilt drive motor is 0.8-3.0W·min / °; Presentation Component The no-load power of the drive system in operation is 5-15W; Presentation Component In time The absolute value of the azimuth adjustment rate; Presentation Component In time The absolute value of the tilt adjustment rate; Represents an indicator function that takes the value 1 when the azimuth or inclination is being adjusted (the rate is not zero), and 0 otherwise.

[0037] Step 4.3, apply the dynamic programming algorithm to solve the optimal angle adjustment trajectory considering the mechanical constraints; During the solution process, the following constraints need to be considered: Component angle adjustment speed limit: , ; Component angle range limitations: , ; Adjacent components are spaced apart to avoid blocking each other: ; in, Represents photovoltaic modules The absolute value of the azimuth adjustment rate, Represents photovoltaic modules The absolute value of the tilt adjustment rate, Indicates the maximum adjustment rate allowed for the azimuth angle. Indicates the maximum adjustment rate allowed for the tilt angle. 、 Represent the minimum and maximum allowed values ​​of the azimuth, 、 Represent the minimum and maximum allowable values ​​of the inclination angle, Indicates that the angle setting and Next, photovoltaic modules and photovoltaic panels The minimum distance between Indicates the minimum safe distance allowed between adjacent components.

[0038] The dynamic programming solution process uses the time window Discretized into time points, calculate the optimal angle state for each time point, and finally generate a complete angle adjustment trajectory , azimuth trajectory The initial value of is based on the solar azimuth prediction data, and is corrected in combination with the distribution characteristics of ambient reflected light. The panoramic light intensity distribution map obtained by the light field perception module determines the direction of the main light source, and then the ambient light contribution model is used to calculate the direction of the main light source. Calculate the light energy contribution weights in each direction and finally determine the optimal azimuth trajectory. The final angle optimization formula is: ; in, represents the time window for prediction, represents the number of discretized time points, Represents photovoltaic modules In time The optimal azimuth angle, Represents photovoltaic modules In time The optimal inclination angle, Indicates the complete angle adjustment trajectory, Indicates the azimuth and inclination parameters that make the objective function reach the maximum value. Indicates that in the time window The cumulative power generation when the component runs according to the angle trajectory, represents the energy consumption function during the angle adjustment process, Indicates the absolute value of the azimuth adjustment rate, Indicates the absolute value of the tilt adjustment rate, Indicates the maximum adjustment rate allowed for the azimuth angle, which is 3-5° / min. Indicates the maximum adjustment rate allowed for the inclination angle, which is 2-3° / min. 、 Respectively represent the minimum and maximum allowable values ​​of the azimuth angle, usually ranging from 0 to 360 degrees, 、 Respectively represent the minimum and maximum allowable values ​​of the inclination angle, usually ranging from 0-90°, Indicates that at a given angle setting, the component and components The minimum distance between Indicates the minimum safe distance allowed between adjacent components to avoid mutual shading, which is 0.5-1.5m. For all different component pairs , all need to satisfy the minimum distance constraint.

[0039] The adaptive angle optimization algorithm of this embodiment is an innovative algorithm, and its specific implementation includes: State space construction: Discretize the angle adjustment trajectory into A grid of states, where and are the discretization levels of azimuth and inclination, respectively (set to 36 and 18); Adaptive time step adjustment: Dynamically adjust the time step according to the predicted rate of light change, using a smaller time step (minimum 1 minute) during periods of drastic light changes and a larger step (maximum 10 minutes) during stable periods; Reward function construction: Comprehensively consider the three factors of predicted power generation, adjusted energy consumption and mechanical stress to form a weighted reward function : Mechanical stress; in, represents the weighted reward function, Derived from the illumination prediction model and component power generation characteristic curve, The energy consumption of angle adjustment is derived from the power characteristics of the drive motor and the angle change rate. The mechanical stress is the mechanical stress index caused by angle adjustment, which is derived from the structural characteristics of the component bracket and the frequency and amplitude of the angle change. 、 、 They represent the weight coefficients of considering predicted power generation, adjusted energy consumption and mechanical stress, respectively, and are set as 、 、 , which can be adjusted dynamically according to the actual situation of the system.

[0040] Forward dynamic programming: Use forward scanning to construct the value function and record the optimal predecessor state for each state. The specific implementation process is as follows: Defining the value function Indicates that at time step In state , the maximum cumulative reward from the current moment to the end moment; For each time step , calculated in order from the initial time to the end time; For each state , calculate its value function: ; in, Indicates time In state Instant rewards, Indicates status The set of all reachable successor states of ; Simultaneously record the optimal precursor state map , used for subsequent path backtracking; Pruning technology is applied in the calculation process to eliminate state transitions that violate the angle adjustment speed constraints and component spacing constraints.

[0041] Reverse path extraction: backtracking from the terminal state to obtain the complete optimal angle trajectory; Step 4.4: Optimize the module tilt angle to maximize energy capture in a specific band based on changes in spectral characteristics; Based on the illumination prediction results, fine-tune the module angle according to the dominant spectral characteristics of different time periods (such as dawn and dusk): ; in, Indicates that after considering the spectral characteristics, the photovoltaic module In time Adjust the castor angle, Represents the photovoltaic components calculated by the dynamic programming algorithm In time The optimal reference inclination angle, Indicates time The dominant wavelength, that is, the wavelength with the highest proportion in the spectral energy distribution, Indicates the dominant wavelength Fine-tune the angle.

[0042] Finally, the optimal angle adjustment instruction sequence for each component in the future time window is output , providing a decision basis for partition collaborative control, among which, Represents photovoltaic modules At the time point The optimal azimuth after adjustment, Represents photovoltaic modules At the time point The optimal tilt angle after adjustment, Indicates inclusion A complete angle adjustment instruction sequence at each time point.

[0043] The partition collaborative control module implements the partition collaborative control system, performs differentiated angle adjustments on PV panels according to angle optimization instructions, and forms an overall optimized control solution; It includes the following sub-steps: Step 5.1, partition the system based on the similarity of lighting characteristics; Assembling components in large photovoltaic systems Apply a clustering algorithm where Represents a collection of photovoltaic modules, Indicates the total number of PV panels; Partitioning based on lighting prediction characteristics and similarity of ambient light contribution: , contains a collection of PV modules with similar illumination characteristics, where 、 、 Respectively represent 、 、 partitions, Indicates the number of regions.

[0044] Partition similarity uses illumination feature vector Calculation, the vector is composed of photovoltaic components The direct light intensity, scattered light intensity, reflected light contribution and other characteristics are obtained by stitching: ; in, Represents photovoltaic modules The lighting characteristic vector, Represents photovoltaic modules and The similarity of lighting characteristics between Represents a vector and The cosine value between and Represents vectors and The mold length.

[0045] when When the photovoltaic module and can be divided into the same area, where Represents the similarity threshold. When the similarity of two components is greater than or equal to the threshold, they are considered to have sufficiently similar lighting characteristics and can be divided into the same area.

[0046] Step 5.2: Build a partition-level angle control model to achieve coordinated adjustment of components within the region; For each partition , establish a unified regional control model: ; in, Represents a partition In time The optimal azimuth and inclination angles, Represents photovoltaic modules The weight coefficient of Indicates finding the azimuth that makes the following expression reach the maximum value and inclination combination, Indicates partition All photovoltaic modules in Perform the summation, Represents photovoltaic modules In azimuth ,inclination and time The predicted power generation under these conditions.

[0047] The angle command obtained by the regional model will be fine-tuned according to the position of each PV module in the region to generate the final module-level command:

[0048] in, Represents photovoltaic modules In time The final azimuth setting value, Represents photovoltaic modules In time The final tilt angle setting value, Presentation Component Azimuth fine-tuning parameters are used to avoid mutual occlusion between components. It is calculated through the inter-component occlusion analysis model, and the calculation formula is: , in, Presentation Component The location coordinates of Indicates the component height, Indicates the location and height information of other components in the same area; Presentation Component The tilt angle fine-tuning parameter is used to avoid mutual occlusion between components. Based on ray tracing algorithm and distance matrix between components Generate, where Presentation Component and components The distance between them. When the tilt angle fine adjustment value is Calculate, where is the proportional coefficient, ranging from 0.5 to 2.0 degrees / meter.

[0049] Step 5.3: Apply the distributed actuator collaborative control algorithm to ensure the smoothness and accuracy of angle adjustment; For each actuator (Control photovoltaic panels angle), establish its dynamic response model: ; in, Represents photovoltaic modules The angular acceleration of represents the dynamic response function of the actuator, Indicates that it is sent to the executive agency The control signal, Represents photovoltaic modules The current angle, Represents photovoltaic modules The current angular velocity, Indicates the actuator Load parameters.

[0050] Based on this model, a distributed model predictive controller (MPC) is configured to optimize the control signal sequence : ; in, Indicates the control signal Perform minimization optimization, Expressing hope for the future The prediction time domain of each time point is summed up. Indicates at a point in time Photovoltaic panels The prediction angle, Indicates at a point in time Photovoltaic panels The target reference angle, represents the square of the angular tracking error, represents the weight coefficient of the squared angular velocity term, represents the square of the angular velocity, represents the weight coefficient of the square term of the control signal, Represents the square of the control signal, which is used to evaluate the control energy consumption, Indicates inclusion A complete sequence of control signals at each time point.

[0051] Step 5.4: Implement inter-regional collaborative optimization to improve overall system performance; For adjacent areas and , establish a coordinated scheduling mechanism to avoid energy fluctuations during angle adjustment: ; in, Indicates that the system is at time The total output power, Indicates that the system is at time The total output power, Indicates the time interval between two adjacent power measurements. Indicates the absolute value of the system power change, Indicates the maximum power fluctuation threshold allowed by the system.

[0052] By staggering the control sequence between regions, a smooth transition of system power is ensured: ; in, Indicates area The execution timing of the area is the time point when the angle adjustment starts. Calculated by the system scheduling algorithm, the calculation formula is: ; Ensure that the angle adjustment time of each area is staggered to avoid drastic fluctuations in system power. Indicates the base time, which is the initial time point when the system starts to perform angle adjustment. Determined by the system main controller based on the predicted light change time, usually set as the advance amount before the predicted light change , the calculation formula is ,in, is the predicted time point of illumination change, The system preset advance adjustment time is 5-15 minutes; Indicates area The timing offset is used to stagger the adjustment time of different areas. Based on the regional power size and the maximum power fluctuation threshold allowed by the system The calculation formula is: ,in, For the region The total power, is the time offset coefficient corresponding to unit power, which is 0.5-2 seconds / kW.

[0053] Finally, the actuator control instruction sequence of each photovoltaic module is output , realizing differentiated angle optimization control of the entire system.

[0054] Real-world application examples of this implementation: This application scenario involves a commercial building complex in a city in East China, with a total roof area of ​​approximately 8,000 square meters. A scalable BIPV system based on intelligent control is installed. The system has a total installed capacity of 960 kWp and comprises 1,200 monocrystalline silicon photovoltaic modules, each with a power output of 800 watts and a size of 2.2 m x 1.1 m. The region's climate is characterized by distinct four seasons, with frequent thunderstorms in the summer and haze in the winter. Annual sunshine hours are approximately 2,100.

[0055] The building's surroundings are complex, with high-rise buildings to the east and south, an open area to the west, and a cluster of mid- and low-rise buildings to the north. The rooftops are irregularly shaped, with some obstructions (such as equipment rooms and cooling towers). This complex environment results in varying light conditions for the photovoltaic panels in different areas, making it difficult to achieve optimal power generation efficiency with traditional, unified angle control strategies.

[0056] The system uses the intelligent control-based angle optimization method for BIPV rooftops proposed in this implementation. The rooftop PV system is divided into eight control zones, each containing 150 PV modules. The overall system layout is shown in Table 1.

[0057] Table 1: Layout of photovoltaic systems on commercial building rooftops;

[0058] Panoramic light field perception system implementation: To achieve comprehensive perception of complex lighting environments, the system deploys a panoramic light field perception system according to the following configuration: A 360° panoramic light field sensor array is installed at the center of the roof. It comprises 24 evenly distributed light intensity sensors, 24 angle sensors, and 8 spectrum analyzers, forming a uniformly distributed spherical sensing network. Furthermore, 32 ground-based light sensors are installed around the edges and in the center of the roof to collect reflected and scattered light data. The system is also equipped with four panoramic cameras, mounted at the four corners of the roof, to capture sky image sequences. The sensor deployment is shown in Table 2.

[0059] Table 2: Sensor deployment of panoramic light field perception system;

[0060] The system accesses satellite cloud images and weather forecast data every 10 minutes through an API, including information such as cloud cover, precipitation probability, temperature, and wind speed for the next six hours. The data synchronization and spatial registration processing unit integrates these data types based on timestamps and spatial coordinates to generate a multidimensional lighting environment dataset.

[0061] Under typical cloudy weather conditions, the average light intensity data collected by the system at different times of the day are shown in Table 3.

[0062] Table 3: Light intensity data at different times of typical cloudy weather (unit: W / m²);

[0063] As can be seen from Table 3, under typical cloudy weather conditions, the cumulative contribution of scattered light and reflected light accounts for about 25%-35% of the total light intensity, and the dominant spectral range at different times also has obvious differences, which proves the necessity of comprehensively collecting a variety of light information.

[0064] Implementation of multimodal data fusion prediction algorithm: The multimodal data fusion prediction algorithm used in this example is optimized for the unique environment of the commercial building rooftop. The multi-stream convolutional long short-term memory network uses four parallel processing branches to process light field sensor data, ground illumination data, sky image sequences, and meteorological satellite data.

[0065] For the sky image sequence processing branch, since the region is prone to thunderstorms in summer, the system enhances its ability to identify rapidly changing clouds. The network structure is as follows: Input layer: Receives a sequence of sky images with a resolution of 256×256 and a time step of 12 (i.e., 6 minutes of data); The first convolutional LSTM layer: 32 5×5 convolution kernels with a stride of 1; The second convolutional LSTM layer has 48 3×3 convolution kernels with a stride of 2. The third convolutional LSTM layer: 64 3×3 convolution kernels with a stride of 1; Feature extraction layer: After global average pooling, a 256-dimensional feature vector is obtained; The adaptive attention fusion module is adjusted to the seasonal characteristics of the region, especially considering the special lighting conditions of winter haze weather. The system's weight distribution of each data source under different meteorological conditions is shown in Table 4.

[0066] Table 4: Data source weight distribution under different meteorological conditions;

[0067] In actual operation, the system's prediction accuracy for lighting changes within the next 60 minutes reaches 93.2%, and its prediction accuracy for lighting changes in the next 60-120 minutes reaches 86.5%. In particular, for sudden changes in lighting caused by cloud movement, the system can predict 5-8 minutes in advance, providing sufficient response time for angle adjustment.

[0068] Adaptive angle optimization algorithm implementation: In this example, the adaptive angle optimization algorithm was specifically configured for the commercial building's roof structure and environmental characteristics. The system discretized the angle adjustment trajectory for each control area into a grid state space with 36 levels of azimuth (0-360°, 10° steps) and 18 levels of inclination (0-90°, 5° steps).

[0069] During operation, the system dynamically adjusts the time step according to the predicted rate of change of light. A smaller time step (2 minutes) is used during the two periods of rapid light change, 9:00-11:00 in the morning and 15:00-17:00 in the afternoon, and a larger time step (8 minutes) is used during the period of relatively stable light, 12:00-14:00 in the afternoon.

[0070] The effect of forward-looking angle optimization is particularly significant in cloudy and windy weather conditions. Table 5 shows the system's effect on PV panel angle optimization in Region 4 (south) under typical cloudy conditions.

[0071] Table 5: Optimization effect of module angles in region 4 under typical cloudy weather conditions;

[0072] Through proactive adjustments, the system optimizes the angle before lighting conditions change, avoiding the energy loss caused by delayed adjustments in traditional passive response systems. Furthermore, the system optimizes the module tilt based on the changing spectral characteristics of different times of day, for example increasing the tilt in the morning and evening to better capture energy in specific wavelengths (600-800nm).

[0073] Partition collaborative control implementation: This example implements zoned collaborative control based on illumination similarity for the eight control zones of the commercial building's rooftop photovoltaic system. The system first calculates each zone's illumination characteristic vector based on illumination predictions, including direct light intensity, diffuse light intensity, and reflected light contribution. It then clusters the zones using cosine similarity.

[0074] The similarity threshold is set to 0.85, and the system eventually divides the eight areas into three collaborative control groups: the east group (areas 1, 2, and 3), the south group (area 4), and the west group (areas 5, 6, 7, and 8). Each collaborative control group implements a unified basic angle control strategy and then makes fine adjustments based on the specific location.

[0075] In order to ensure the smoothness of angle adjustment, the system reasonably arranges the adjustment sequence of the three control groups, as shown in Table 6.

[0076] Table 6: Timing arrangements of angle adjustments for different control groups;

[0077] This staggered timing design ensures a smooth transition of overall output power during large-scale power adjustments, preventing sudden power surges from impacting the grid. The system's maximum power fluctuation threshold is set at 5% of the total installed capacity, or 48kW. In actual operation, the maximum power fluctuation remains below 38kW.

[0078] For each actuator, the system applies a distributed model predictive controller for precise control, achieving smooth and accurate angle adjustment while taking into account mechanical constraints. In the controller's optimization objectives, the position accuracy weight λ is set to 0.6, the velocity smoothness weight μ is set to 0.3, and the control cost weight is set to 0.1, achieving a balance between high accuracy and smooth angle adjustment.

[0079] Technical effect verification: This implementation method has been validated for commercial rooftop photovoltaic systems over a 12-month period, collecting sufficient data to verify its technical effectiveness. This effectiveness is primarily demonstrated in two key areas: comprehensive improvements in light capture efficiency and proactive control.

[0080] Comprehensively improve light energy capture efficiency: This system significantly improves light capture efficiency by comprehensively utilizing direct, scattered, and reflected light. Based on year-round operational data, this system's average annual power generation efficiency is 16.1% higher than that of conventional single-light source tracking systems. This improvement is even more pronounced under adverse weather conditions (cloudy, rainy, and foggy), with improvements exceeding 20%. Especially under cloudy conditions, this system achieves a power generation efficiency of 77.8% (relative to the theoretical maximum efficiency), compared to 48.7% for traditional fixed-angle systems and 56.2% for conventional single-light source tracking systems, an improvement of 21.6%. This demonstrates that this system's ability to comprehensively utilize direct, scattered, and reflected light significantly improves light capture efficiency.

[0081] Proactive control effects: Table 7 shows the comparison between this system and conventional real-time response systems under conditions of rapid lighting changes.

[0082] Table 7: Comparison of system response under conditions of rapid light changes;

[0083] Data shows that the system can respond to changes in sunlight an average of 6 minutes in advance, while conventional systems lag behind by an average of 3.2 minutes. This results in a 19.1% increase in power generation efficiency under conditions of rapidly changing sunlight. This demonstrates the importance of proactive control in photovoltaic systems.

[0084] In summary, this embodiment has demonstrated significant technical effects in practical applications, especially in comprehensively improving the light energy capture efficiency and achieving forward-looking control, which fully verifies the practicality and advancement of the present invention.

[0085] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A BIPV rooftop power system based on intelligent control, characterized in that: include: The panoramic light field perception module uses a panoramic light field perception system to collect multi-dimensional lighting environment data including direct light, scattered light, and reflected light, and generates real-time light field distribution information; The multimodal data fusion prediction module uses a multimodal data fusion prediction algorithm to process real-time light field distribution information and form a prediction model for light changes in the next 120 minutes. The ambient light contribution analysis module uses a ray tracing algorithm combined with a light change prediction model to analyze the ambient light contribution and generate a light gain model that reflects the impact of different light sources on components in each area. Adaptive angle optimization module, which executes the adaptive angle optimization algorithm to calculate the optimal angle adjustment trajectory of the photovoltaic module based on the light gain model and outputs the angle optimization instruction; The partition collaborative control module implements the partition collaborative control system, performs differentiated angle adjustments on photovoltaic modules according to angle optimization instructions, and forms an overall optimization control solution.

2. The intelligent control-based BIPV rooftop power system according to claim 1 is characterized in that: The panoramic light field perception system includes: The 360° panoramic light field sensor array is configured to collect all-round light intensity, angle and spectral distribution information; A network of ground-based light sensors deployed to collect reflected and scattered light data; Panoramic camera, satellite cloud image receiver and meteorological data collector, used to collect sky image sequences, satellite cloud images and meteorological forecast data; The data synchronization and spatial registration processing unit is used to integrate various types of data into a multi-dimensional lighting environment dataset according to time and space coordinates.

3. The intelligent control-based BIPV rooftop power system according to claim 1 is characterized in that: The multimodal data fusion prediction algorithm includes: Multi-stream convolutional long short-term memory networks, including spherical convolutional long short-term memory units, two-dimensional convolutional long short-term memory networks, deep residual convolutional long short-term memory networks, and multi-layer perceptron and long short-term memory network hybrid models, are used to process different types of input data; Adaptive attention fusion module, including feature transformation layer, weather condition encoder, time period encoder, conditional attention calculation unit, multi-head attention mechanism and feature fusion layer; The multi-scale temporal prediction network, including short-term prediction model, medium-term prediction model and long-term prediction model, is used to generate illumination prediction results for different time windows.

4. The intelligent control-based BIPV rooftop power system according to claim 1 is characterized in that: The ray tracing algorithm includes: Construct a 3D model of the environment surrounding the PV system, including spatial information and material properties of buildings, terrain, and other objects; Apply the bidirectional reflectance distribution function model to calculate the reflective characteristics of the environment surface to light; Implement the ray tracing process of multiple reflections and calculate the influence weights of different light sources on components in each area; A dynamic light gain model is constructed to quantify the contribution of environmental factors to photovoltaic power generation at different times.

5. The intelligent control-based BIPV rooftop power system according to claim 1 is characterized in that: The adaptive angle optimization algorithm includes: Establish a model for the relationship between photovoltaic module angle and power generation efficiency, describing the relationship between incident light intensity, incident angle, and module photoelectric conversion efficiency; Construct a forward-looking optimization objective function to maximize the cumulative power generation within the forecast period; A dynamic programming algorithm is applied to solve the optimal angle adjustment trajectory under the constraints of component angle adjustment speed limit, angle range limit and adjacent component spacing setting; Based on the changes in spectral characteristics, the component tilt angle is optimized to maximize energy capture in specific bands.

6. The intelligent control-based BIPV rooftop power system according to claim 1 is characterized in that: The partition collaborative control system includes: A clustering algorithm for system partitioning based on the similarity of light characteristics is used to divide the component set in the photovoltaic system into multiple areas with similar light characteristics; A partition-level angle control model to achieve coordinated adjustment of components within a region; Distributed actuator collaborative control algorithm to ensure the smoothness and accuracy of angle adjustment; Inter-regional collaborative optimization mechanism is used to ensure smooth transition of system power by controlling timing staggering.

7. The intelligent control-based BIPV rooftop power system according to claim 3 is characterized in that: The structure of the multi-stream convolutional long short-term memory network includes: Input layer, used to receive sensory data of different modalities; Multiple parallel convolutional LSTM layers are used to simultaneously capture the spatial features and temporal dependencies of different modal sensory data; Feature extraction layer, used to obtain a fixed-dimensional feature vector through global average pooling; The output layer is used to output the features extracted from each modal data.

8. The intelligent control-based BIPV rooftop power system according to claim 5 is characterized in that: The implementation of the adaptive angle optimization algorithm includes: State space construction, discretizing the angle adjustment trajectory into a grid state space with discretization levels; Adaptive time step adjustment, dynamically adjusting the time step according to the predicted rate of illumination change; The reward function is constructed by comprehensively considering multiple factors such as predicted power generation, adjusted energy consumption and mechanical stress; Forward dynamic programming, constructing a value function and recording the optimal predecessor state for each state; Reverse path extraction, backtracking from the terminal state to obtain the complete optimal angle trajectory.

9. The intelligent control-based BIPV rooftop power system according to claim 1 is characterized in that: When the multimodal data fusion prediction algorithm processes the real-time light field distribution information, different weights are assigned to each data source according to different weather conditions and time periods: On sunny days, a higher weight is given to direct light data; In overcast or cloudy weather, higher weights are given to scattered light and satellite cloud image data; During sunrise and sunset, higher weights are given to spectral data.

10. A computer-readable storage medium, characterized in that It is used to store computer-readable instructions, and when the computer-readable instructions are read by a computer, it can run a liftable BIPV photovoltaic rooftop power system based on intelligent control as described in any one of claims 1-9.

Images