Photovoltaic street light dynamic linkage control system based on multi-source environmental perception

Abstract

This application provides a dynamic linkage control system for photovoltaic streetlights based on multi-source environmental perception, relating to the field of intelligent lighting control technology. It identifies targets based on multi-source heterogeneous environmental data from photovoltaic streetlights to obtain dynamic environmental characteristics. Energy status assessment is performed based on the state data of the photovoltaic energy storage system and meteorological forecast data to generate energy status assessment data. A multi-source perception map is constructed based on the environmental dynamic characteristics and energy status assessment data. Then, strategy calculation is performed on the multi-source perception map to generate a lighting grouping and scheduling strategy. A linkage twin model of the photovoltaic streetlights is constructed. The lighting grouping and scheduling strategy is input into the linkage twin model for virtual simulation and security verification to obtain a security simulation result. The security simulation result is decomposed into a sequence of individual control commands for each photovoltaic streetlight. This application can achieve dynamic optimization of photovoltaic streetlights in complex environments based on multi-source environmental perception, energy status assessment, and digital twin simulation verification.

Description

Technical Field

[0001] This application relates to the field of intelligent lighting control technology, and more specifically, to a dynamic linkage control system for photovoltaic streetlights based on multi-source environmental perception. Background Technology

[0002] As an important component of smart city infrastructure, intelligent lighting control systems have undergone several stages of technological evolution, from traditional timed control and light-controlled switches to remote single-lamp control based on the Internet of Things. With the rapid development of sensor technology, wireless communication technology, and artificial intelligence algorithms, modern intelligent lighting systems have gradually acquired the ability to dynamically adjust lighting intensity according to ambient light, traffic density, and pedestrian distribution, thereby improving road lighting safety while achieving energy conservation and emission reduction.

[0003] By deploying light sensors, microwave radar, and video acquisition units on streetlight poles to perceive environmental conditions and generating dimming commands based on preset control logic or lightweight algorithms, the photovoltaic power generation system is deeply integrated with intelligent lighting control. This achieves two goals: firstly, by utilizing photovoltaic energy storage systems to provide clean energy for streetlights, reducing the load on the urban power grid; and secondly, by using multi-source sensor fusion to achieve synchronized lighting of streetlight groups according to dynamic changes in traffic flow. However, existing dynamic linkage control technologies for photovoltaic streetlights based on multi-source environmental perception still suffer from limitations. They only monitor the current state of charge of the photovoltaic energy storage system, lacking a mechanism for predicting future energy supply capacity. They execute control strategies directly after generation, lacking prior safety verification, and often employ rule engines or single-objective optimization algorithms, making it difficult to achieve dynamic balance among multiple interdependent objectives such as traffic safety, pedestrian experience, and system energy consumption. Therefore, how to achieve dynamic optimization of photovoltaic streetlights in complex environments based on multi-source environmental perception through energy status assessment and digital twin simulation verification remains a challenge for the industry. Summary of the Invention

[0004] This application provides a dynamic linkage control system for photovoltaic streetlights based on multi-source environmental perception. It can realize the dynamic optimization of photovoltaic streetlights in complex environments through energy status assessment and digital twin simulation verification based on multi-source environmental perception.

[0005] This application provides a dynamic linkage control system for photovoltaic streetlights based on multi-source environmental perception, the dynamic linkage control system comprising: The data acquisition module is used to acquire multi-source heterogeneous environmental data of photovoltaic streetlights and status data of photovoltaic energy storage systems; The situation assessment module is used to identify targets based on the multi-source heterogeneous environmental data, obtain environmental dynamic characteristics, and perform energy situation assessment based on the state data and meteorological forecast data to generate energy situation assessment data. The strategy perception module is used to construct a multi-source perception map based on the environmental dynamics and the energy situation assessment data, input the multi-source perception map into a multi-objective deep reinforcement learning model for strategy calculation, and generate a lighting grouping and scheduling strategy. The simulation and verification module is used to construct a linkage twin model of photovoltaic streetlights based on the multi-source heterogeneous environmental data, input the lighting grouping and scheduling strategy into the linkage twin model for virtual simulation and security verification, and obtain the security simulation results. The instruction decomposition module is used to decompose the safety simulation results into a sequence of single-lamp control instructions for photovoltaic streetlights.

[0006] In this embodiment, multi-source heterogeneous environmental data of the photovoltaic street light is acquired through a sensor array, and status data of the photovoltaic energy storage system is acquired through a current and voltage monitoring unit.

[0007] In this embodiment, target identification based on the multi-source heterogeneous environmental data to obtain dynamic environmental features specifically includes: The multi-source heterogeneous environmental data is spatiotemporally aligned and multimodal fused to obtain a fused perception tensor; The lightweight convolutional neural network pre-trained with the fused perceptual tensor input is used for target detection and trajectory prediction to obtain the environmental category label, location coordinates and motion speed. Based on the category label, the location coordinates and the movement speed, a spatiotemporal dynamic feature matrix is ​​constructed, and then the spatiotemporal dynamic feature matrix is ​​visualized as environmental dynamic features.

[0008] In this embodiment, the energy situation assessment is performed based on the state data and meteorological forecast data, and the energy situation assessment data is generated specifically including: Obtain meteorological forecast data through the meteorological information center; The state data and the meteorological forecast data are input into a long short-term memory neural network for time series modeling and state prediction to obtain the remaining power prediction curve of the photovoltaic street light. The remaining power prediction curve is subjected to hierarchical quantification to obtain energy status assessment data.

[0009] In this embodiment, constructing a multi-source sensing map based on the environmental dynamics and the energy state assessment data specifically includes: The environmental dynamic characteristics are mapped to targets to obtain a target spatial distribution map; The energy situation assessment data is mapped to levels to obtain a spatial distribution map of the energy situation. By cascading channels and aligning features of the target spatial distribution map and the energy situation spatial distribution map, a multi-source sensing map is obtained.

[0010] In this embodiment, inputting the multi-source sensing map into a multi-target deep reinforcement learning model for policy calculation to generate a lighting grouping scheduling policy specifically includes: The multi-source perception map is input into the feature extraction network of the multi-objective deep reinforcement learning model to extract features and obtain the state feature vector. The state feature vector is input into the policy network of the multi-objective deep reinforcement learning model for policy evaluation to obtain the brightness adjustment coefficient and the switch scheduling state. The lighting grouping and scheduling strategy is determined based on the brightness adjustment coefficient and the switch scheduling status.

[0011] In this embodiment, a linkage twin model of photovoltaic streetlights is constructed based on the multi-source heterogeneous environmental data using data visualization and digital twin technology.

[0012] In this embodiment, the lighting grouping and scheduling strategy is input into the linked twin model for virtual simulation and security verification, and the security simulation results specifically include: The linked twin model is driven to perform virtual lighting based on the lighting grouping and scheduling strategy. During the virtual lighting process, illuminance distribution data and visibility distance data are collected. Threshold verification was performed on the illuminance distribution data and the visible distance data respectively to obtain the safety simulation results.

[0013] In this embodiment, decomposing the safety simulation results into a sequence of single-lamp control commands for photovoltaic streetlights specifically includes: The safety simulation results are analyzed element by element to obtain the time-brightness discrete sequence of the photovoltaic street light; The discrete time-brightness sequence is encapsulated into a standard instruction frame for an addressable lighting interface, thereby obtaining a single-lamp control instruction sequence.

[0014] In this embodiment, the multi-objective deep reinforcement learning model is a neural network model constructed based on the dual-delay deep deterministic policy gradient algorithm.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The process involves acquiring multi-source heterogeneous environmental data for photovoltaic streetlights and state data for photovoltaic energy storage systems; performing target identification based on the multi-source heterogeneous environmental data to obtain dynamic environmental characteristics; conducting energy situation assessment based on the state data and meteorological forecast data to generate energy situation assessment data; constructing a multi-source sensing map based on the dynamic environmental characteristics and the energy situation assessment data; inputting the multi-source sensing map into a multi-target deep reinforcement learning model for strategy calculation to generate a lighting grouping and scheduling strategy; constructing a linkage twin model for photovoltaic streetlights based on the multi-source heterogeneous environmental data; inputting the lighting grouping and scheduling strategy into the linkage twin model for virtual simulation and safety verification to obtain a safety simulation result; and decomposing the safety simulation result into a sequence of single-lamp control commands for the photovoltaic streetlights.

[0016] Therefore, this application enables dynamic optimization of photovoltaic streetlights in complex environments. Firstly, by acquiring multi-source heterogeneous environmental data of the photovoltaic streetlights and state data of the photovoltaic energy storage system, the physical basis of the road dynamic environment and its own energy state is perceived from multiple dimensions, providing data support for subsequent refined scheduling. Secondly, based on the multi-source heterogeneous environmental data, target identification is performed to obtain environmental dynamic characteristics, and energy situation assessment data is generated based on state data and meteorological forecast data, achieving real-time perception of traffic conditions and forward-looking prediction of future energy supply capacity. Thirdly, a multi-source perception map is constructed based on the environmental dynamic characteristics and energy situation assessment data, and this map is input into a multi-objective deep reinforcement learning model for policy implementation. The simplified calculation generates a lighting grouping scheduling strategy, which facilitates the unified representation of heterogeneous spatiotemporal information and its input into an intelligent decision-making model for multi-objective joint optimization. Then, a linkage twin model of photovoltaic streetlights is constructed based on multi-source heterogeneous environmental data. The lighting grouping scheduling strategy is input into this linkage twin model for virtual simulation and security verification to obtain a safe simulation result. A digital security verification barrier is built before the strategy is deployed to the physical world, which helps avoid security risks caused by strategy deviations. Finally, the safety simulation result is decomposed into a sequence of individual control commands for the photovoltaic streetlights, enabling precise conversion from the verified global scheduling strategy to executable physical commands. This improves the operational safety and energy efficiency of the photovoltaic streetlight system in complex dynamic environments.

[0017] In summary, the technical solution adopted in this application can realize the dynamic optimization of photovoltaic streetlights in complex environments through energy status assessment and digital twin simulation verification based on multi-source environmental perception. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a module structure diagram of the photovoltaic street light dynamic linkage control system based on multi-source environmental perception provided in this application; Figure 2 This is a flowchart illustrating the process of determining dynamic environmental characteristics in some embodiments of this application; Figure 3 This is a schematic diagram of the process for determining a multi-source sensing map in some embodiments of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides a dynamic linkage control system for photovoltaic streetlights based on multi-source environmental perception. Its core is to acquire multi-source heterogeneous environmental data of the photovoltaic streetlights and state data of the photovoltaic energy storage system; perform target identification based on the multi-source heterogeneous environmental data to obtain dynamic environmental characteristics; perform energy situation assessment based on the state data and meteorological forecast data to generate energy situation assessment data; construct a multi-source perception map based on the dynamic environmental characteristics and the energy situation assessment data; input the multi-source perception map into a multi-target deep reinforcement learning model for strategy calculation to generate a lighting grouping scheduling strategy; construct a linkage twin model of the photovoltaic streetlights based on the multi-source heterogeneous environmental data; input the lighting grouping scheduling strategy into the linkage twin model for virtual simulation and safety verification to obtain a safety simulation result; and decompose the safety simulation result into a sequence of individual control commands for each photovoltaic streetlight.

[0022] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 As shown in the figure, this is a modular structure diagram of the photovoltaic street light dynamic linkage control system based on multi-source environmental perception provided in this application. The dynamic linkage control system includes: a data acquisition module 100, a situation assessment module 200, a strategy perception module 300, a deduction and verification module 400, and an instruction decomposition module 500, which are described below: The data acquisition module 100 is used to acquire multi-source heterogeneous environmental data of photovoltaic streetlights and status data of photovoltaic energy storage systems.

[0023] In this specific implementation, a sensor array is used to acquire multi-source heterogeneous environmental data of the photovoltaic streetlights, and a current and voltage monitoring unit is used to acquire status data of the photovoltaic energy storage system. The sensor array includes a light sensor array, a millimeter-wave radar sensor, and a video acquisition unit. The light sensor array is used to collect light intensity data of the environment where the photovoltaic streetlights are located. The millimeter-wave radar sensor is used to collect traffic flow, vehicle speed, and vehicle position data on the road. The video acquisition unit is used to collect pedestrian distribution and abnormal event image data on the road. The current and voltage monitoring unit refers to a monitoring module installed at the connection between the output end of the photovoltaic energy storage system and the battery pack. It includes a current monitoring unit and a voltage monitoring unit. The current monitoring unit is used to collect charging and discharging current data of the photovoltaic energy storage system in real time, and the voltage monitoring unit is used to collect terminal voltage data of the photovoltaic energy storage system in real time.

[0024] It should be noted that the multi-source heterogeneous environmental data of photovoltaic streetlights refers to multi-dimensional heterogeneous information reflecting the physical environment around the photovoltaic streetlights, collected by different types of sensors, including illuminance data, traffic flow data, pedestrian data, and meteorological data. This is beneficial for providing a comprehensive environmental perception basis for subsequent target identification and energy status assessment. The status data of the photovoltaic energy storage system refers to electrical parameter data reflecting the current energy status and operational health of the photovoltaic cells, including battery state of charge, charging and discharging current, terminal voltage, and battery temperature. This is beneficial for providing accurate current energy status input for energy status assessment. In addition, the photovoltaic energy storage system in this application refers to an integrated system composed of photovoltaic panels, energy storage battery packs, charge and discharge controllers, and energy management units, used to convert solar energy into electrical energy and store it, providing photovoltaic streetlights with a clean energy supply independent of the power grid.

[0025] The situation assessment module 200 is used to identify targets based on the multi-source heterogeneous environmental data, obtain environmental dynamic characteristics, and perform energy situation assessment based on the state data and meteorological forecast data to generate energy situation assessment data.

[0026] Preferred, Reference Figure 2 As shown in the figure, this is a flowchart illustrating the process of determining environmental dynamics in this embodiment. In this embodiment, target identification based on the multi-source heterogeneous environmental data to obtain environmental dynamics can be achieved through the following steps: First, in step S21, the multi-source heterogeneous environmental data is spatiotemporally aligned and multimodal fused to obtain a fused sensing tensor; Then, in step S22, the lightweight convolutional neural network pre-trained with the fused perceptual tensor input is used for target detection and trajectory prediction to obtain the environmental category label, location coordinates and motion speed. Finally, in step S23, a spatiotemporal dynamic feature matrix is ​​constructed based on the category label, the location coordinates, and the motion speed, and then the spatiotemporal dynamic feature matrix is ​​visualized as environmental dynamic features.

[0027] It should be noted that, in this embodiment, the pre-trained lightweight convolutional neural network refers to a lightweight convolutional neural network obtained after pre-training on a road scene dataset. The road scene dataset contains road image data and annotation information under different lighting conditions, weather conditions, and traffic densities. The annotation information includes: category labels, location bounding boxes, and direction of motion. In the actual training process, the image data in the road scene dataset can be input into the lightweight convolutional neural network. The target category prediction value, location prediction value, and motion speed prediction value output by the network are calculated through forward propagation. Then, the error between the network output and the annotation information is calculated through the loss function. The network weights are then updated through the backpropagation algorithm. The training is iterated until the loss function converges to obtain the pre-trained lightweight convolutional neural network.

[0028] In practice, firstly, the timestamps of multi-source heterogeneous environmental data are uniformly calibrated using time synchronization technology to eliminate time delays caused by differences in sampling frequencies between different types of sensors, resulting in time-aligned multi-source heterogeneous environmental data. Then, millimeter-wave radar point cloud data is projected from the radar coordinate system to the video image coordinate system through spatial coordinate transformation. Illumination intensity data is added as an additional information layer and fused with the aligned radar point cloud projection and video image along the channel dimension, resulting in a multi-channel 3D data tensor containing illumination, radar point cloud density, and image pixel features. This multi-channel 3D data tensor is used as the fusion perception tensor. Next, the fusion perception tensor is input into a pre-trained lightweight deep convolutional neural network. High-dimensional semantic features are extracted through the network's multi-layer convolutional operations. The network's target detection branch outputs the category label and bounding box position coordinates of each detected target in the image. Simultaneously, the network's trajectory prediction branch calculates the instantaneous velocity of the target based on positional changes in the same target across multiple consecutive frames. The category label, position coordinates, and velocity are used as the target recognition result. Finally, the category labels, location coordinates, and movement speeds of all targets in the target recognition results are organized into a matrix according to time order and spatial location. The rows of the matrix correspond to different targets, and the columns of the matrix correspond to the category label values, location coordinate values, and movement speed values, respectively, thus constructing a spatiotemporal dynamic feature matrix. Then, the pseudo-color mapping technology is used to map each element value in the spatiotemporal dynamic feature matrix to a pixel color value, generating a two-dimensional image that can intuitively reflect the spatiotemporal distribution pattern of dynamic traffic targets. This two-dimensional image is used as the environmental dynamic feature.

[0029] It should be noted that the fusion perception tensor in this application refers to the retention of the original feature information of illumination, radar point cloud and image pixels; the spatiotemporal dynamic feature matrix refers to a structured data table that organizes the category, position and velocity information of each target in the target recognition result according to the time and space dimensions, which can transform discrete detection results into a regular data form; the environmental dynamic feature refers to the two-dimensional image generated by mapping the spatiotemporal dynamic feature matrix through visualization, which presents the category distribution, spatial position and motion state of traffic targets in the road environment in a visual way.

[0030] In this embodiment, energy situation assessment is performed based on the state data and meteorological forecast data. The energy situation assessment data can be generated in the following manner: Obtain meteorological forecast data through the meteorological information center; The state data and the meteorological forecast data are input into a long short-term memory neural network for time series modeling and state prediction to obtain the remaining power prediction curve of the photovoltaic street light. The remaining power prediction curve is subjected to hierarchical quantification to obtain energy status assessment data.

[0031] In practical implementation, firstly, a program interface is written to send a request to the meteorological data service platform to obtain the hourly solar irradiance prediction sequence and the hourly ambient temperature prediction sequence for the area where the photovoltaic streetlights are located. These two sequences are then used together as meteorological forecast data. Next, the status data collected by the current and voltage monitoring units are aligned and concatenated with the hourly solar irradiance prediction sequence and the hourly ambient temperature prediction sequence from the meteorological forecast data, constructing a multi-dimensional time series input long short-term memory neural network. The long short-term dependencies in the time series are modeled through the input gate, forget gate, and output gate of this long short-term memory neural network, allowing the output of the remaining power prediction value for each time point within a preset time period. First, the forget gate calculates a forgetting coefficient between 0 and 1 based on the input data of the current time step and the hidden state passed from the previous time step. That is, the input data of the current time step and the hidden state of the previous time step are concatenated into a combined vector, and this combined vector is input into the fully connected layer of the forget gate. After the linear transformation of the fully connected layer, the output is compressed to between 0 and 1 by the sigmoid activation function to obtain the forgetting coefficient. Each value in the forgetting coefficient corresponds to the proportion of the positional information retained in the state of the previous time step. A value close to 1 indicates that a large amount of positional information is retained, and a value close to 0 indicates that a large amount of positional information is discarded. Then, the input gate calculates an input coefficient between 0 and 1 and a candidate memory vector based on the input data of the current time step and the hidden state of the previous time step. That is, the combined vector formed by concatenating the input data of the current time step and the hidden state of the previous time step is passed to the two fully connected layers of the input gate. One fully connected layer outputs the input coefficient after passing through the sigmoid activation function, and the other fully connected layer outputs the candidate memory vector after passing through the hyperbolic tangent activation function. Each value in the input coefficient corresponds to the proportion of information added to the corresponding position in the candidate memory vector. A value close to 1 indicates that a large amount of candidate information at that position is added to the memory, and a value close to 0 indicates that a small amount of candidate information at that position is added to the memory. The forgetting coefficient is multiplied element-wise with the state of the previous time step to obtain the retained historical information. The input coefficient is multiplied element-wise with the candidate memory vector to obtain the new information to be added. Finally, the retained historical information is added element-wise with the new information to be added to obtain the updated state of the current time step. Finally, the output gate calculates an output coefficient between 0 and 1 based on the input data of the current time step and the hidden state of the previous time step. That is, the combined vector formed by concatenating the input data of the current time step and the hidden state of the previous time step is input into the fully connected layer of the output gate, and the output coefficient is solved by the sigmoid activation function; the updated state of the current time step is compressed to between -1 and 1 by the hyperbolic tangent activation function, and then multiplied element-wise with the output coefficient to obtain the hidden state of the current time step. This hidden state is the output result of the current time step, and is carried to the next time step as the hidden state of the previous time step of the next time step.

[0032] Preferably, the preset time period can be set to 30 minutes, which facilitates timely processing and scheduling of photovoltaic streetlights. Then, the remaining power prediction values ​​for all time points are connected in chronological order, and a continuous photovoltaic streetlight remaining power prediction curve is generated using a visualization library. Finally, the remaining power prediction values ​​for each time point in the remaining power prediction curve are compared. An energy level label corresponding to each time point is determined by a pre-defined energy level threshold. All energy level labels for all time points are arranged in chronological order, and the resulting sequence is used as energy status assessment data. The pre-defined energy level threshold can be set based on the actual service life of the photovoltaic streetlights; for example, for new batteries with a service life of less than one year: High energy level corresponds to the range [80%, 100%], with an energy level label of 3; medium energy level corresponds to the range [50%, 80%), with an energy level label of 2; low energy level corresponds to the range [20%, 50%), with an energy level label of 1; and very low energy level corresponds to the range [0%, 20%), with an energy level label of 0. For batteries aged between three and five years: high energy level corresponds to the range [70%, 100%], energy level label is 3; medium energy level corresponds to the range [40%, 70%), energy level label is 2; low energy level corresponds to the range [15%, 40%), energy level label is 1; very low energy level corresponds to the range [0%, 15%), energy level label is 0. For batteries that have been in service for more than five years and are nearing the end of their service life: High energy level corresponds to the range [60%, 100%], with an energy level label of 3; Medium energy level corresponds to the range [30%, 60%), with an energy level label of 2; Low energy level corresponds to the range [10%, 30%], with an energy level label of 1; Very low energy level corresponds to the range [0%, 10%], with an energy level label of 0.

[0033] By matching the predicted remaining power value at each time point in the remaining power prediction curve with the energy level division interval corresponding to the above-mentioned service life, the energy level label corresponding to each time point in the remaining power prediction curve can be determined. The energy level labels of all time points are arranged in chronological order, and the sorted sequence is used as energy status assessment data.

[0034] It should be noted that the meteorological forecast data in this application refers to numerical information on future weather conditions, including predicted values ​​of solar irradiance and ambient temperature, which are used to estimate the potential energy input of the photovoltaic power generation system; the remaining power prediction curve is a numerical curve that depicts the dynamic evolution of the available electrical energy of the photovoltaic streetlights over a period of time; and the energy status assessment data characterizes the sufficiency of the photovoltaic streetlights' power supply at each moment, which can provide the energy dimension decision input basis for the strategy perception module.

[0035] The strategy perception module 300 is used to construct a multi-source perception map based on the environmental dynamic characteristics and the energy situation assessment data, input the multi-source perception map into a multi-objective deep reinforcement learning model for strategy calculation, and generate a lighting grouping and scheduling strategy.

[0036] Preferred, Reference Figure 3 As shown in the figure, this is a flowchart illustrating the process of determining the multi-source sensing map in this embodiment. In this embodiment, the construction of the multi-source sensing map based on the environmental dynamics and the energy state assessment data can be achieved through the following steps: First, in step S31, the environmental dynamic features are mapped to obtain a target spatial distribution map; Then, in step S32, the energy situation assessment data is mapped to levels to obtain an energy situation spatial distribution map; Finally, in step S33, the target spatial distribution map and the energy situation spatial distribution map are channel-cascaded and feature-aligned to obtain a multi-source sensing map.

[0037] In practice, firstly, the location coordinates of each dynamic traffic target in the environmental dynamic features are transformed from the image pixel coordinate system to the road geographic coordinate system through coordinate transformation. The location of each target is then marked on a pre-constructed blank grid map. Each grid cell in the grid map corresponds to a fixed area in the road geographic space. During marking, the corresponding grid cell is determined based on the geographic coordinates of the target, and a numerical value representing the target's presence and category information is assigned to this grid cell. After traversing all targets, a two-dimensional grid image is generated, which serves as the target spatial distribution map. Next, each energy level label in the energy situation assessment data is associated with the actual location of the photovoltaic streetlights in geographic space through geographic indexing. On another blank grid map, the corresponding grid cell is determined based on the geographic coordinates of each photovoltaic streetlight, and the grid cell is assigned the corresponding energy level label value. After assigning values ​​to all streetlights, a two-dimensional grid image is generated, which serves as the energy situation spatial distribution map. Finally, the target spatial distribution map and the energy situation spatial distribution map are geometrically aligned through spatial registration to ensure that the grid cells of the same geographical location in the two images correspond one-to-one in space. Then, the aligned target spatial distribution map and energy situation spatial distribution map are superimposed in the channel dimension to generate a three-dimensional data cube containing information from both channels. This three-dimensional data cube is used as a multi-source sensing map.

[0038] It should be noted that the target spatial distribution map in this application refers to a map that intuitively presents the density distribution and spatial location of various targets on the road in the form of grid cell values; the energy status spatial distribution map refers to a map that intuitively presents the distribution of the energy sufficiency of streetlights in different areas in the form of grid cell values; and the multi-source perception map refers to three-dimensional data containing traffic target distribution information and energy status distribution information, which can provide spatially aligned and feature-rich joint input for subsequent multi-target deep reinforcement learning models.

[0039] In this embodiment, the multi-source sensing map is input into a multi-target deep reinforcement learning model for policy calculation to generate a lighting grouping scheduling policy, which can be done in the following way: The multi-source perception map is input into the feature extraction network of the multi-objective deep reinforcement learning model to extract features and obtain the state feature vector. The state feature vector is input into the policy network of the multi-objective deep reinforcement learning model for policy evaluation to obtain the brightness adjustment coefficient and the switch scheduling state. The lighting grouping and scheduling strategy is determined based on the brightness adjustment coefficient and the switch scheduling status.

[0040] It should be noted that in this embodiment, the multi-objective deep reinforcement learning model is a neural network model constructed based on the dual-delay deep deterministic policy gradient algorithm. Specifically, a model structure containing four neural networks is first constructed: an online policy network, a target policy network, an online value network, and a target value network. The online policy network generates action policies based on the input state feature vector; the target policy network provides stable target values ​​for updating the value network; the online value network evaluates the merits of the current action policy; and the target value network calculates the target value to stabilize the training process. The collaborative connection of these four networks is as follows: the actions generated by the online policy network are used for environmental interaction and generate empirical data. After random sampling from the empirical data, the target policy network and the target value network jointly calculate the target value. The online value network calculates the current value, updates the online value network using the difference between the two, and then updates the online policy network using the policy gradient. Simultaneously, the parameters of the online network are slowly copied to the target network. Through this collaborative training and delayed update mechanism, the dual-delay deep deterministic policy... The gradient algorithm can learn stable and efficient control strategies in a continuous action space, providing intelligent decision support for the dynamic optimization of photovoltaic streetlights in complex environments. Furthermore, the reward function of this dual-delay deep deterministic strategy gradient algorithm is obtained by summing a safety reward term and an energy consumption reward term, used to evaluate the merits of each action strategy. The safety reward term is the reward value calculated during virtual simulation based on illuminance distribution data and visibility distance data; specifically, it is set to 1 when the illuminance values ​​of all sampling points are not lower than the safe lighting threshold and the visibility distance values ​​of all traffic targets are not lower than the safe line-of-sight threshold, otherwise it is set to -1. The energy consumption reward term is the reward value calculated during virtual simulation based on the brightness adjustment coefficients and switching sequence of all photovoltaic streetlights in the lighting grouping and scheduling strategy. It can be obtained by multiplying the average brightness adjustment coefficient values ​​of all photovoltaic streetlights by 0.5, adding the proportion of all photovoltaic streetlights in the off state to the total number of streetlights multiplied by 0.5, and then taking the negative value of the sum.

[0041] In practice, firstly, the multi-source sensing map is input into a multi-objective deep reinforcement learning model. The multi-source sensing map is convolved through multiple convolutional layers to extract local feature maps at different spatial scales. Then, the local feature maps are downsampled through pooling layers to reduce spatial resolution and retain the main features. Finally, the multi-dimensional feature map obtained after downsampling is flattened and mapped into a one-dimensional numerical sequence through a fully connected layer. This one-dimensional numerical sequence is used as the state feature vector. Then, the state feature vector is input into the policy network of the multi-objective deep reinforcement learning model. The state feature vector is transformed layer by layer nonlinearly through multiple fully connected layers in the policy network. In the last fully connected layer, the brightness adjustment coefficient and on / off state values ​​corresponding to each photovoltaic street light are output respectively. For the original value of the brightness adjustment coefficient, the linear activation function of the fully connected layer is used to map the original value of the brightness adjustment coefficient to the interval between the minimum and maximum brightness values. The value transformed by the linear activation function is used as the brightness adjustment coefficient. For the original value of the on / off state, the normalized exponential function of the fully connected layer is used to convert the original values ​​of the on and off states corresponding to each photovoltaic street light into on and off state probability values. The on / off state is set to 1 when the probability of the photovoltaic street light being on is greater than the probability of being off, indicating the on state. The on / off state is set to 0 when the probability of the photovoltaic street light being on is less than the probability of being off, indicating the off state. Finally, the brightness adjustment coefficients of all photovoltaic streetlights are arranged in the order of their geographical location numbers to obtain a brightness adjustment coefficient sequence. The switching statuses of all photovoltaic streetlights are arranged in the same order to obtain a switching status sequence. The brightness adjustment coefficient sequence and the switching status sequence are combined into a matrix data structure containing two rows of data. This matrix data structure is used as the lighting grouping scheduling strategy.

[0042] It should be noted that the state feature vector in this application refers to the key information that integrates traffic target distribution and energy situation distribution, which can provide a basis for decision-making for the strategy network; the brightness adjustment coefficient refers to the continuous value generated by the strategy network for each photovoltaic street light. This value corresponds to the proportion of the street light's luminous intensity in the next control cycle. A brightness adjustment coefficient of 1 indicates that the street light is emitting light at maximum power, a brightness adjustment coefficient of 0.5 indicates that the street light is emitting light at half power, and a brightness adjustment coefficient of 0 indicates that the street light is not emitting light at all. Through this coefficient, the luminous intensity of the street light can be dynamically adjusted under different traffic flow and different ambient light conditions. For example, when there are no cars or people passing by. The brightness adjustment coefficient is reduced to achieve energy saving, and increased to ensure safety when there are heavy traffic and pedestrians. The switch scheduling state refers to the discrete state value generated by the strategy network for each photovoltaic street light. This value indicates whether the street light should be in the on or off state in the next control cycle. It is used to realize the grouping on and off control of street lights. If the switch scheduling state of a photovoltaic street light is off (0), then the brightness adjustment coefficient of that photovoltaic street light is also 0. The lighting grouping scheduling strategy refers to the decision set composed of the brightness adjustment coefficients and switch scheduling states of all photovoltaic street lights. This set describes the cooperative working mode of photovoltaic street lights in a structured form.

[0043] The simulation and verification module 400 is used to construct a linkage twin model of photovoltaic streetlights based on the multi-source heterogeneous environmental data, input the lighting grouping and scheduling strategy into the linkage twin model for virtual simulation and security verification, and obtain the security simulation results.

[0044] It should be noted that in this embodiment, a linked twin model of the photovoltaic street light is constructed based on the multi-source heterogeneous environmental data using data visualization and digital twin technology. Specifically, firstly, a panoramic scan of the target road section is performed using a laser scanning device deployed on the photovoltaic street light tower, collecting road geographic information data, three-dimensional structural data of the street light tower, and outline data of surrounding buildings. The collected point cloud data is then imported into three-dimensional reconstruction software for surface mesh reconstruction and texture mapping to generate a static three-dimensional scene model. Then, the photovoltaic energy storage system status data collected by the current and voltage monitoring unit, the environmental data collected by the sensor array, and the dynamic traffic target data obtained from target recognition are connected to the three-dimensional real-time rendering engine through a real-time data interface. This drives the virtual street light model in the static three-dimensional scene model to update its illumination status synchronously according to the real-time status data, and drives the virtual traffic target model to update its motion trajectory synchronously according to the real-time position coordinates, generating a dynamic three-dimensional scene that moves synchronously with the physical world in real time. This dynamic three-dimensional scene serves as the linked twin model of the photovoltaic street light. Preferably, data visualization can use 3D reconstruction software to construct 3D scenes, which is beneficial for efficiently processing point cloud data and generating realistic static scene models; digital twin technology can use a 3D real-time rendering engine for dynamic data access and model driving, which is beneficial for achieving high-fidelity virtual-real synchronization and real-time visualization interaction. Other existing technologies can also be used in other embodiments, which are not limited here.

[0045] In this embodiment, the lighting grouping and scheduling strategy is input into the linked twin model for virtual simulation and security verification. The security simulation results can be obtained in the following manner: The linked twin model is driven to perform virtual lighting based on the lighting grouping and scheduling strategy. During the virtual lighting process, illuminance distribution data and visibility distance data are collected. Threshold verification was performed on the illuminance distribution data and the visible distance data respectively to obtain the safety simulation results.

[0046] In specific implementation, firstly, the brightness adjustment coefficient sequence and the switch scheduling state sequence in the lighting grouping and scheduling strategy are assigned to the corresponding virtual street light models in the linkage twin model according to the geographical location number of the photovoltaic street light. This drives the virtual street light models to adjust the luminous intensity according to the assigned brightness adjustment coefficients and control the opening and closing of the luminous units according to the switch scheduling state. A complete control cycle of virtual lighting process is executed in the 3D real-time rendering engine. During the execution of the virtual lighting process, the illuminance values ​​of each sampling point in the key areas of the road are collected by the virtual illuminance sensor array preset in the linkage twin model. The illuminance values ​​of all sampling points are organized into an illuminance distribution grid according to their spatial location, and this illuminance distribution grid is used as the illuminance distribution data. At the same time, the visible distance values ​​between the target and the surrounding environment are collected by the virtual vision sensor preset in the linkage twin model, simulating the driver's perspective at the location of the dynamic traffic target. The visible distance values ​​of all traffic targets are organized into a visible distance sequence according to the target number, and this visible distance sequence is used as the visible distance data. Then, each illuminance value in the illuminance distribution data is compared with a preset safe lighting threshold to determine if the illuminance values ​​at all sampling points are not lower than the safe lighting threshold. If the illuminance values ​​at all sampling points are not lower than the safe lighting threshold, an illuminance verification pass label with a value of 1 is generated; if the illuminance value at any sampling point is lower than the safe lighting threshold, an illuminance verification fail label with a value of 0 is generated. Simultaneously, each visible distance value in the visible distance data is compared with a preset safe sight distance threshold to determine if the visible distance values ​​of all traffic targets are not lower than the safe sight distance threshold. If the visible distance values ​​of all traffic targets are not lower than the safe sight distance threshold, a sight distance verification pass label with a value of 1 is generated; if the visible distance value of any traffic target is lower than the safe sight distance threshold, a sight distance verification fail label with a value of 0 is generated. The safe lighting threshold can be set based on the actual spacing between photovoltaic streetlights, for example, taking half the streetlight spacing as the minimum illuminance requirement for the road surface; the safe sight distance threshold can be set based on the safe stopping sight distance corresponding to the road's design speed. Finally, the labels for passing or failing illuminance verification are combined with the labels for passing or failing line-of-sight verification into a tuple containing the two verification results, and this tuple is used as the safety simulation result.

[0047] In addition, if the tuple contains a tag indicating that the illuminance verification failed or the line-of-sight verification failed, it means that there is a security risk in the lighting grouping and scheduling strategy. The lighting grouping and scheduling strategy can be discarded, and the strategy regeneration mechanism can be triggered. That is, a recalculation request is sent to the strategy perception module, which then regenerates the lighting grouping and scheduling strategy based on the latest multi-source perception map. The strategy is then input into the simulation and verification module for virtual simulation and security verification until a tuple containing two verified tags is obtained. If the number of consecutive regenerations exceeds the preset maximum number of attempts threshold (for example, the preset maximum number of attempts threshold is 3), all photovoltaic streetlights can be turned on at maximum brightness, and an alarm message is sent through the communication terminal interface.

[0048] It should be noted that the illuminance distribution data in this application refers to the spatial distribution information of illuminance intensity in key road areas during the virtual lighting process. The illuminance levels at different locations are presented in a grid format and used to evaluate whether the lighting grouping and scheduling strategy meets road lighting safety standards. The visibility distance data refers to the identifiable distance information of the environment around the photovoltaic streetlights during the virtual lighting process. The visibility conditions of different traffic targets are presented in a numerical sequence format and used to evaluate whether the lighting grouping and scheduling strategy ensures safe driving visibility. The safety simulation results refer to the results that clearly indicate whether the current scheduling strategy has passed the safety verification and can provide a decision-making basis for the subsequent instruction decomposition module.

[0049] The instruction decomposition module 500 is used to decompose the safety simulation results into a sequence of single-lamp control instructions for photovoltaic streetlights.

[0050] In this embodiment, the safety simulation results can be decomposed into a sequence of individual control commands for photovoltaic streetlights in the following manner: The safety simulation results are analyzed element by element to obtain the time-brightness discrete sequence of the photovoltaic street light; The discrete time-brightness sequence is encapsulated into a standard instruction frame for an addressable lighting interface, thereby obtaining a single-lamp control instruction sequence.

[0051] In specific implementation, firstly, the binary pairs in the safety simulation results are judged. If the binary pair contains the illuminance verification pass label and the line-of-sight verification pass label, then the brightness adjustment coefficient sequence and the switch scheduling state sequence in the lighting grouping scheduling strategy are extracted. Each brightness adjustment coefficient value in the brightness adjustment coefficient sequence is associated with the corresponding photovoltaic street light number, and each switch state value in the switch scheduling state sequence is associated with the corresponding photovoltaic street light number. Then, according to the preset control cycle time resolution, the brightness adjustment coefficient value and switch state value of each photovoltaic street light in each time slice are combined into a time-state pair. The time-state pairs in all time slices are arranged in chronological order to obtain the time-brightness discrete sequence corresponding to each photovoltaic street light. Then, each time-state pair in the discrete time-brightness sequence corresponding to each photovoltaic street light is encoded according to the data format of the digital addressable lighting interface protocol to generate a binary data frame containing a start code, address code, operation code, brightness data code, and check code. The address code is filled with the physical address number of the current photovoltaic street light, the brightness data code is filled with the brightness adjustment coefficient value in the time-state pair, and the operation code is filled with an on or off command according to the switch status value. All binary data frames are arranged in chronological order and street light number order to obtain an ordered set of binary data frames. This ordered set of binary data frames is used as the single-lamp control command sequence.

[0052] It should be noted that the time-brightness discrete sequence in this application refers to a data sequence formed by arranging the brightness adjustment coefficients and on / off states of each photovoltaic street light in each time slice within the control cycle in chronological order. This sequence describes in detail the dynamic working mode of each street light throughout the entire control cycle. The single-lamp control command sequence refers to a set of binary data frames generated by encapsulating the time-brightness discrete sequence according to the digital addressable lighting interface protocol. Each data frame in this set corresponds to a control command for a street light in a specific time slice, which can be directly parsed and executed by the street light controller after being sent through the communication network.

[0053] In summary, the technical solution adopted in this application can realize the dynamic optimization of photovoltaic streetlights in complex environments.

[0054] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0055] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0056] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A dynamic linkage control system for photovoltaic streetlights based on multi-source environmental perception, characterized in that, The dynamic linkage control system includes: The data acquisition module is used to acquire multi-source heterogeneous environmental data of photovoltaic streetlights and status data of photovoltaic energy storage systems; The situation assessment module is used to identify targets based on the multi-source heterogeneous environmental data, obtain environmental dynamic characteristics, and perform energy situation assessment based on the state data and meteorological forecast data to generate energy situation assessment data. The strategy perception module is used to construct a multi-source perception map based on the environmental dynamics and the energy situation assessment data, input the multi-source perception map into a multi-objective deep reinforcement learning model for strategy calculation, and generate a lighting grouping and scheduling strategy. The simulation and verification module is used to construct a linkage twin model of photovoltaic streetlights based on the multi-source heterogeneous environmental data, input the lighting grouping and scheduling strategy into the linkage twin model for virtual simulation and security verification, and obtain the security simulation results. The instruction decomposition module is used to decompose the safety simulation results into a sequence of single-lamp control instructions for photovoltaic streetlights.

2. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, Multi-source heterogeneous environmental data of photovoltaic streetlights are acquired through sensor arrays, and status data of photovoltaic energy storage systems are acquired through current and voltage monitoring units.

3. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, Target identification based on the aforementioned multi-source heterogeneous environmental data yields dynamic environmental features, specifically including: The multi-source heterogeneous environmental data is spatiotemporally aligned and multimodal fused to obtain a fused perception tensor; The lightweight convolutional neural network pre-trained with the fused perceptual tensor input is used for target detection and trajectory prediction to obtain the environmental category label, location coordinates and motion speed. Based on the category label, the location coordinates and the movement speed, a spatiotemporal dynamic feature matrix is ​​constructed, and then the spatiotemporal dynamic feature matrix is ​​visualized as environmental dynamic features.

4. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, Based on the aforementioned state data and meteorological forecast data, an energy situation assessment is performed, and the generated energy situation assessment data specifically includes: Obtain meteorological forecast data through the meteorological information center; The state data and the meteorological forecast data are input into a long short-term memory neural network for time series modeling and state prediction to obtain the remaining power prediction curve of the photovoltaic street light. The remaining power prediction curve is subjected to hierarchical quantification to obtain energy status assessment data.

5. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, Constructing a multi-source sensing map based on the environmental dynamics and energy state assessment data specifically includes: The environmental dynamic characteristics are mapped to targets to obtain a target spatial distribution map; The energy situation assessment data is mapped to levels to obtain a spatial distribution map of the energy situation. By cascading channels and aligning features of the target spatial distribution map and the energy situation spatial distribution map, a multi-source sensing map is obtained.

6. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, The multi-source sensing map is input into a multi-target deep reinforcement learning model for policy calculation, and the generated lighting grouping scheduling policy specifically includes: The multi-source perception map is input into the feature extraction network of the multi-objective deep reinforcement learning model to extract features and obtain the state feature vector. The state feature vector is input into the policy network of the multi-objective deep reinforcement learning model for policy evaluation to obtain the brightness adjustment coefficient and the switch scheduling state. The lighting grouping and scheduling strategy is determined based on the brightness adjustment coefficient and the switch scheduling status.

7. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, A linked twin model of photovoltaic streetlights is constructed based on the multi-source heterogeneous environmental data using data visualization and digital twin technology.

8. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, The lighting grouping and scheduling strategy is input into the linked twin model for virtual simulation and security verification. The specific security simulation results include: The linked twin model is driven to perform virtual lighting based on the lighting grouping and scheduling strategy. During the virtual lighting process, illuminance distribution data and visibility distance data are collected. Threshold verification was performed on the illuminance distribution data and the visible distance data respectively to obtain the safety simulation results.

9. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, The safety simulation results are decomposed into a sequence of individual control commands for photovoltaic streetlights, specifically including: The safety simulation results are analyzed element by element to obtain the time-brightness discrete sequence of the photovoltaic street light; The discrete time-brightness sequence is encapsulated into a standard instruction frame for an addressable lighting interface, thereby obtaining a single-lamp control instruction sequence.

10. The photovoltaic street light dynamic linkage control system based on multi-source environmental perception as described in claim 1, characterized in that, The multi-objective deep reinforcement learning model is a neural network model constructed based on the dual-delay deep deterministic policy gradient algorithm.

Images