Intelligent street lamp adaptive control system based on edge gateway

Abstract

A smart street lamp adaptive control system based on an edge gateway comprises a server side and an edge gateway side which are interconnected through a network. The server side is internally provided with an intelligent street lamp self-adaptive control general model, and the edge gateway side is provided with an intelligent street lamp self-adaptive control sub-model, a lighting demand extraction model, a street lamp control selection module and a street lamp self-adaptive adjustment equation set module; the system carries out feature extraction on environment data and street lamp data through the illumination demand extraction model, after judgment of the control selection module, feature vectors can be sent to the control total model to generate first adjustment vectors, and then second adjustment vectors are obtained through calculation of the sub-models; or an adjustment vector is obtained through direct calculation of the adjustment equation set, and finally a street lamp control instruction is generated. The thought of combining total points is adopted, and the technical problem that in the prior art, a unified control strategy is often adopted, and it is difficult to differentiate and respond to illumination requirements in real time is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of street lamp control and relates to an intelligent street lamp adaptive control system based on an edge gateway. Background Art

[0002] With the in-depth advancement of smart city construction, the intelligentization and energy saving of urban lighting systems have become important development directions. Traditional urban street light control systems mainly use simple time control or light control methods, which cannot be accurately adjusted according to actual environmental requirements, resulting in energy waste and poor lighting effects. At present, the smart street light control system mainly has the following application scenarios and technical difficulties: 1. Lighting requirements in complex urban environments: Modern urban roads are complex and changeable, including commercial areas, residential areas, industrial areas and other functional areas, and the lighting requirements of each area are significantly different. Existing technologies often adopt a unified control strategy, which is difficult to meet differentiated lighting needs.

[0003] 2. Dynamic traffic flow adaptation: Urban traffic flow has significant temporal and spatial distribution characteristics. Existing control systems mostly use preset solutions and cannot respond to traffic flow changes in real time, resulting in uneven distribution of lighting resources. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a smart street lamp adaptive control system based on an edge gateway, which solves the technical problem that the existing technology often adopts a unified control strategy and is difficult to differentiate and respond to lighting needs in real time.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a smart street light adaptive control system based on an edge gateway, characterized in that: it includes a server end and an edge gateway end interconnected by a network; a smart street light adaptive control general model is arranged in the server end, and a smart street light adaptive control sub-model, a lighting demand extraction model, a street light control selection module and a street light adaptive adjustment equation group module are arranged in the edge gateway end; the lighting demand extraction model is used to extract features from environmental data and street light data to obtain feature vectors of each type of data, and after completing the feature vector extraction, all the obtained feature vectors are sent to the smart street light adaptive control general model in the server end; the smart street light adaptive control general model calculates all the received feature vectors to obtain a first street light adjustment vector, and sends it to the edge gateway end; the street light control selection module determines whether to send all the feature vectors to the smart street light adaptive control general model or to the street light adaptive adjustment equation group module; the street light adaptive adjustment equation group module calculates all the feature vectors obtained by the lighting demand extraction model through a preset empirical equation group.

[0006] The smart street light adaptive control sub-model is a fine-tuned Lora model of the smart street light adaptive control overall model, which is used to enhance the adaptability of the adaptive control of street lights connected to the edge gateway; the smart street light adaptive control sub-model calculates the first street light adjustment vector to obtain the second street light adjustment vector for generating street light control instructions.

[0007] The overall model of the smart street lamp adaptive control adopts the Transformer encoder-decoder structure; the sub-model of the smart street lamp adaptive control adopts the lightweight LoRA structure; and the lighting demand extraction model adopts the CNN+LSTM hybrid structure.

[0008] The street light adaptive adjustment equation group includes basic lighting equation, pedestrian flow response equation, vehicle flow response equation, weather compensation equation and energy efficiency optimization equation; the street light adjustment vector includes brightness value, start and stop time, gradient rate and group linkage parameter.

[0009] The basic lighting equation is as follows: ; In the formula, is the basic brightness value, ranging from [0,1]; is the ambient light intensity (lux); is the maximum reference illuminance, which is 10000 lux; It is the 24-hour time (h); is the basic lighting coefficient, taking the value [0.6,0.8]; is the night compensation coefficient, with a value of [0.2, 0.4]; is the time attenuation coefficient, which takes a value of 0.1; is the error term.

[0010] The flow response equation is as follows: ; In the formula, The supplementary value of pedestrian lighting, range [0,0.3]; is the current crowd density (people / m²); is the maximum crowd density, which is 5 people / m²; is the activity coefficient of the population; is the maximum activity coefficient, which takes the value of 1; is the weight coefficient, which takes values ​​of 0.2, 0.1, and 0.15 respectively; is the change rate of human flow density; is the error term.

[0011] The traffic flow response equation is as follows: ; In the formula, Supplementary value for traffic lighting, range [0,0.4]; is the current traffic density (vehicles / km); is the maximum traffic density, which is 200 vehicles / km; is the average vehicle speed (km / h); is the reference speed, which is 60km / h; are weight coefficients, taking values ​​of 0.3 and 0.2 respectively; is the weight of the i-th vehicle model; is the distance attenuation coefficient; is the distance from the i-th vehicle to the street light (m); is the number of vehicle types; is the error term.

[0012] The weather compensation equation is as follows: ; In the formula, is the weather compensation value, ranging from [0,0.2]; is the current visibility (m); is the standard visibility, with a value of 1000m; is the precipitation intensity (mm / h); is the wind speed (m / s); is the relative humidity; is the weight coefficient, which takes values ​​of 0.15, 0.1, and 0.05 respectively; is the humidity influence coefficient, the value is 0.5; is an imaginary unit used to represent the phase effect of wind speed; is the error term.

[0013] The energy efficiency optimization equation is as follows: ; In the formula, is the final adjustment value, ranging from [0,1]; is the current power (W); is the rated power (W); are weight coefficients, taking values ​​of 1.0 and 0.1 respectively; is the kth energy efficiency factor weight; is the deviation value of the kth energy efficiency indicator; is the number of energy efficiency indicators; is the error term.

[0014] The selection criteria of the street light control selection module are: when the network communication is normal and the computing resources are sufficient, the smart street light adaptive control general model + smart street light adaptive control sub-model collaborative approach is used; when the network delay exceeds the threshold or the bandwidth is limited, the street light adaptive adjustment equation group module is used; when an emergency is detected, the local control mode, that is, the street light adaptive adjustment equation group module, is used; at the same time, the edge gateway adjusts the street light through PWM dimming signals and digital switching quantities.

[0015] The main beneficial effects of the present invention are: The "general-specific combination" deep learning control architecture solves the problem of coordinated control of large-scale street lighting systems.

[0016] Applying LoRA technology to edge computing scenarios significantly reduces computing overhead.

[0017] A complete set of empirical equations is designed as downgrade protection, providing a reliable backup mechanism.

[0018] Through the combination of edge computing and artificial intelligence technology, precise control of smart street lamps is achieved; the layered architecture reduces the system's computing load and network bandwidth requirements; through the coordination of multiple control strategies, the system's reliability and adaptability are improved; adaptive adjustment based on real-time data optimizes lighting effects and energy efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0020] Figure 1 It is a system diagram of the present invention.

[0021] Figure 2 This is a comparison chart of lighting control parameters in different areas of the present invention.

[0022] Figure 3 It is a comparison diagram of the 24-hour energy consumption curve of the present invention.

[0023] Figure 4 It is a lighting response curve diagram under different weather conditions of the present invention.

[0024] Figure 5 It is a thermal diagram of system performance indicators of the present invention. DETAILED DESCRIPTION

[0025] like Figure 1~Figure 5In the invention, a smart street light adaptive control system based on edge gateway is characterized in that: it comprises a server and an edge gateway interconnected through a network; a smart street light adaptive control master model is arranged in the server, and a smart street light adaptive control sub-model, a lighting demand extraction model, a street light control selection module and a street light adaptive adjustment equation group module are arranged in the edge gateway; the lighting demand extraction model is used to extract features from environmental data and street light data to obtain feature vectors of each data, and after the feature vector extraction is completed, all the obtained feature vectors are sent to the smart street light adaptive control master model in the server; the smart street light adaptive control master model calculates all the received feature vectors to obtain a first street light adjustment vector, and sends it to the edge gateway; the street light control selection module determines whether to send all the feature vectors to the smart street light adaptive control master model or to the street light adaptive adjustment equation group module; the street light adaptive adjustment equation group module calculates all the feature vectors obtained by the lighting demand extraction model through a preset empirical equation group.

[0026] In the preferred scheme, the smart street light adaptive control sub-model is a fine-tuned Lora model of the smart street light adaptive control overall model, which is used to enhance the adaptability of the adaptive control of the street lights connected to the edge gateway; the smart street light adaptive control sub-model calculates the first street light adjustment vector to obtain the second adjustment vector of the street light, which is used to generate the street light control instruction.

[0027] In the preferred solution, the overall model of the smart street lamp adaptive control adopts a Transformer encoder-decoder structure; the sub-model of the smart street lamp adaptive control adopts a lightweight LoRA structure; and the lighting demand extraction model adopts a CNN+LSTM hybrid structure.

[0028] In the preferred scheme, the street light adaptive adjustment equation group includes basic lighting equation, pedestrian flow response equation, vehicle flow response equation, weather compensation equation and energy efficiency optimization equation; the street light adjustment vector includes brightness value, start and stop time, gradient rate and group linkage parameters.

[0029] Preferably, the environmental data includes pedestrian flow, vehicle flow, ambient light intensity, weather conditions, and time period within the range of the street lights to which the edge gateway is connected; the street light data includes the voltage, current, power, operating time, and fault status of each smart street light.

[0030] Preferably, the overall model of the smart street lamp adaptive control adopts a Transformer encoder-decoder structure, specifically a multi-head self-attention mechanism to process multimodal input features, perform feature fusion through a deep residual network, and finally output a control vector through a fully connected layer.

[0031] Preferably, the smart street lamp adaptive control sub-model adopts a lightweight LoRA structure, specifically, a low-rank adaptation layer is added on the basis of the original Transformer, and the model is fine-tuned by updating the weights of the adaptation layer to reduce the computational complexity.

[0032] Preferably, the lighting demand extraction model adopts a CNN+LSTM hybrid structure, specifically using CNN to extract spatial features and LSTM to process temporal features, so as to achieve feature extraction and fusion of multimodal data.

[0033] Preferably, the steps for obtaining the training data set of the overall model of smart street lamp adaptive control specifically include: collecting historical street lamp operation data in multiple areas of the city; organizing environmental data and control effect data; data cleaning and standardization processing; and constructing input feature vectors and label pairs.

[0034] Preferably, the steps for obtaining the training data set of the smart street lamp adaptive control sub-model specifically include: collecting real-time data within the coverage area of ​​the local edge gateway; recording the control effect feedback of the overall model; constructing a local feature data set; and establishing a fine-tuning training set.

[0035] In the preferred scheme, the selection criteria of the street light control selection module are: when the network communication is normal and the computing resources are sufficient, the smart street light adaptive control overall model + smart street light adaptive control sub-model collaborative method is used; when the network delay exceeds the threshold or the bandwidth is limited, the street light adaptive adjustment equation group module is used; when an emergency is detected, the local control mode is used, that is, the street light adaptive adjustment equation group module; at the same time, the edge gateway adjusts the street light through PWM dimming signals and digital switching quantities.

[0036] Embodiment 1, The steps to establish the training data set for the lighting demand extraction model mainly include four key links: The first is the collection of multi-source heterogeneous data. It is necessary to comprehensively collect street light sensor data, including basic parameters such as voltage, current, power, operating time, fault status, etc. At the same time, collect environmental monitoring data including environmental factors such as pedestrian flow, vehicle flow, light intensity, weather conditions, etc., and record time period information in detail, such as distinguishing between weekdays / holidays, morning and evening peak / off-peak periods, and other time series characteristics.

[0037] Then, systematic data preprocessing is carried out, including detection and removal of outliers, filling and interpolation of missing data, alignment and synchronization of time series data, and standardization and normalization of data.

[0038] Secondly, in-depth feature engineering is carried out, including feature extraction based on time windows, calculation of statistical features such as mean, variance, peak, etc., construction of combined features, and analysis and screening of correlations between features.

[0039] The last step is to construct label data. It is necessary to combine expert experience to label the optimal lighting intensity, consider the evaluation indicators of historical control effects, establish a complete multi-dimensional evaluation system, and finally generate high-quality training sample labels to ensure the scientificity and representativeness of the data set.

[0040] The steps for lighting requirement extraction model training are a step-by-step process: First, to initialize the model, you need to reasonably configure the CNN network structure and parameters, set the appropriate number of LSTM layers and hidden units, scientifically initialize the model weights, and define appropriate loss functions and optimizers.

[0041] The second is the data loading and batch processing stage, including the reasonable division of training sets and validation sets, the design of data generators, the construction of data loading pipelines, and the implementation of batch data preprocessing.

[0042] Next is the training process control, which requires setting appropriate training rounds and batch sizes, implementing a dynamic adjustment mechanism for the learning rate, adding an effective early stopping mechanism, and configuring regular saving of model checkpoints.

[0043] Again, model training and optimization, including calculating losses through forward propagation, updating parameters through backpropagation, evaluating performance on the validation set, and dynamically adjusting hyperparameters based on the evaluation results to optimize the training effect.

[0044] The last step is model evaluation and deployment. It is necessary to conduct performance evaluation on an independent test set, compress and quantize the trained model, generate a model version suitable for deployment, and establish an online effect monitoring mechanism to ensure that the model can run stably and reliably in actual applications and continue to provide accurate lighting demand prediction support for the smart street lighting system. The entire training process focuses on the balance between efficiency and accuracy, and through standardized process design and strict quality control, ensures that the final model has good generalization ability and practical value.

[0045] Embodiment 2, The steps for obtaining the training data set of the overall model of smart street lamp adaptive control include: collecting historical street lamp operation data in multiple areas of the city; organizing environmental data and control effect data; data cleaning and standardization; and constructing input feature vectors and label pairs.

[0046] The training steps of the overall model of smart street lamp adaptive control include: Pre-training stage: self-supervised learning using large-scale data; Fine-tuning phase: supervised learning using scenario-specific data; Verification and optimization phase: Use the validation set to evaluate and optimize the model; Deployment testing phase: testing and improvement in the actual environment.

[0047] The steps for obtaining the training data set of the smart street lamp adaptive control sub-model include: collecting real-time data within the coverage area of ​​the local edge gateway; recording the control effect feedback of the overall model; constructing a local feature data set; and establishing a fine-tuning training set.

[0048] The training steps of the smart street lamp adaptive control sub-model include: loading the pre-trained total model weights; freezing the backbone network parameters; only updating the LoRA adaptation layer parameters; and using the fine-tuning training set for incremental learning.

[0049] The street lamp adaptive adjustment equation group includes basic lighting equation, pedestrian flow response equation, vehicle flow response equation, weather compensation equation and energy efficiency optimization equation.

[0050] The basic lighting equation is used to calculate the basic lighting requirements. The input is time and ambient light intensity, and the output is the basic brightness value. The crowd response equation is used to calculate the lighting demand corresponding to the crowd density. The input is the crowd flow and activity characteristics, and the output is the crowd lighting supplement value. The traffic response equation is used to calculate the lighting demand corresponding to the traffic density. The input is the traffic volume and speed, and the output is the traffic lighting supplement value. The weather compensation equation is used to calculate the lighting compensation in bad weather. The input is weather conditions and visibility, and the output is the weather compensation value. The energy efficiency optimization equation is used to balance lighting effect and energy consumption. The input is the total lighting demand and power load, and the output is the final adjustment value.

[0051] Embodiment 3, The specific expression of the street light adaptive adjustment equation group is as follows: 1. The basic lighting equation is specifically expressed as follows: ; In the formula, is the basic brightness value, ranging from [0,1]; is the ambient light intensity (lux); is the maximum reference illuminance, which is 10000 lux; It is the 24-hour time (h); is the basic lighting coefficient, taking the value [0.6,0.8]; is the night compensation coefficient, with a value of [0.2, 0.4]; is the time attenuation coefficient, which takes a value of 0.1; is the error term.

[0052] 2. The crowd flow response equation is specifically expressed as follows: ; In the formula, The supplementary value of pedestrian lighting, range [0,0.3]; is the current crowd density (people / m²); is the maximum crowd density, which is 5 people / m²; is the activity coefficient of the population; is the maximum activity coefficient, which takes the value of 1; is the weight coefficient, which takes values ​​of 0.2, 0.1, and 0.15 respectively; is the change rate of human flow density; is the error term.

[0053] 3. The traffic flow response equation is specifically expressed as follows: ; In the formula, Supplementary value for traffic lighting, range [0,0.4]; is the current traffic density (vehicles / km); is the maximum traffic density, which is 200 vehicles / km; is the average vehicle speed (km / h); is the reference speed, which is 60km / h; are weight coefficients, taking values ​​of 0.3 and 0.2 respectively; is the weight of the i-th vehicle model; is the distance attenuation coefficient; is the distance from the i-th vehicle to the street light (m); is the number of vehicle types; is the error term.

[0054] 4. The weather compensation equation is specifically expressed as follows: ; In the formula, is the weather compensation value, ranging from [0,0.2]; is the current visibility (m); is the standard visibility, with a value of 1000m; is the precipitation intensity (mm / h); is the wind speed (m / s); is the relative humidity; is the weight coefficient, which takes values ​​of 0.15, 0.1, and 0.05 respectively; is the humidity influence coefficient, the value is 0.5; is an imaginary unit used to represent the phase effect of wind speed; is the error term.

[0055] 5. The energy efficiency optimization equation is specifically expressed as follows: ; In the formula, is the final adjustment value, ranging from [0,1]; is the current power (W); is the rated power (W); are weight coefficients, taking values ​​of 1.0 and 0.1 respectively; is the kth energy efficiency factor weight; is the deviation value of the kth energy efficiency indicator; is the number of energy efficiency indicators; is the error term.

[0056] Parameter acquisition method description: 1. Directly measured by light sensor; 2. Obtain it by following these steps: Step 1: Use an infrared sensor array to detect the heat source distribution of the human body; Step 2: People counting using computer vision algorithms: ; In the formula, Pixel The number of people detection function at ; is the detection area (m²); is the number of sampling points.

[0057] 3. Obtain it by following these steps: Step 1: Use geomagnetic sensors to detect vehicle passing information; Step 2: Calculate the number of vehicles passing per unit time: ; In the formula, For the moment Vehicle counts; is the road length (km); is the statistical time window (h).

[0058] 4. Obtained through direct measurement by visibility sensors; 5. Obtained through the power parameter measurement module: ; In the formula, is the effective voltage (V); is the effective current (A); is the power factor.

[0059] 1. Principle explanation of basic lighting equation: (1) Adoption The term represents the inverse relationship between ambient light and lighting requirements. The stronger the ambient light, the less lighting is required, which is consistent with the principle of natural light compensation. (2) Adoption The term represents the periodic change of 24 hours a day, and the sine function can smoothly express the alternation of day and night; (3) Introduction The term is used as night compensation, and the exponential decay reflects the adaptation process of the human eye to darkness, which is derived from the Weber-Fechner law; (4) The multiplication of the various factors reflects the coupling effect between the factors rather than a simple superposition.

[0060] 2. Principle explanation of the flow response equation: (1) Adoption Normalize the crowd density of each item to ensure that the output is within a reasonable range; (2) Introduction The item reflects the changing trend of the flow of people, which is based on the principle of predictive lighting control; (3) Adoption The structure is not directly added to ensure that there is still basic lighting when the change rate of the flow of people is zero; (4) Activity item The use of a similar structure reflects the incremental effect of crowd activities on lighting needs.

[0061] 3. Principle explanation of traffic flow response equation: (1) Adoption The term represents the nonlinear effect of traffic density, and the square relationship reflects the need for rapid response under congestion conditions; (2) Speed ​​​​term Indicates the correction of lighting requirements due to vehicle speed. Fast driving requires better lighting conditions. (3) Introduction The term represents the spatial distribution impact of different types of vehicles, and the exponential decay is based on the law of light intensity decay; (4) The multiplication of various factors reflects the comprehensive effect of traffic flow characteristics, which originates from the flow-density-speed relationship theory in traffic engineering.

[0062] 4. Principle explanation of weather compensation equation: (1) Adoption The term represents the inverse relationship between visibility and compensation, which is based on the principle of atmospheric optics; (2) Introducing plural terms To deal with the influence of wind speed, the imaginary number reflects the phase modulation effect of wind speed on the light distribution; (3) Adoption The term represents the exponential decay effect of humidity, which is derived from the theory of light scattering by the air medium; (4) The precipitation term is linearly superposed , rather than an exponential relationship, which is based on the measured effects of precipitation on visibility.

[0063] 5. Explanation of the principle of energy efficiency optimization equation: (1) Adoption The function limits the final output to no more than 1, ensuring that the control amount is within the physically achievable range; (2) Power term A linear inhibition relationship is adopted to reflect the energy-saving constraint; (3) Introduction The term represents the multiplicative effect of multiple energy efficiency indicators, so that a significant deviation in any one indicator will have a significant impact; (4) Compensation amount The constraints are unified by adding and superposing, which reflects the comprehensive balance of multi-objective optimization.

[0064] Embodiment 4, The server is equipped with a general model for adaptive control of smart street lights, which uses the encoder-decoder structure of the converter in the deep learning architecture. The model contains a multi-head self-attention mechanism for processing multimodal input features, a deep residual network for feature fusion, and a fully connected layer to output a control vector. In the multi-head self-attention mechanism, the number of attention heads is set to 8, and the dimension of each attention head is 64, which can fully capture the correlation between different features. The deep residual network adopts a 6-layer structure, each layer contains two 3×3 convolutional layers and a skip connection, and the activation function uses a rectified linear unit. The fully connected layer adopts a three-layer structure with dimensions of 512, 256, and 128, respectively, and the final output dimension matches the control vector dimension.

[0065] The edge gateway is equipped with multiple functional modules, the core of which is the smart street light adaptive control sub-model. This sub-model adopts a lightweight low-rank adaptive structure and adds a low-rank adaptive layer on the basis of the original converter. The rank of the low-rank adaptive layer is set to 16, which can significantly reduce the computational complexity while maintaining the model performance. The adaptation layer is implemented by matrix decomposition, which decomposes the original weight matrix into the product of two low-rank matrices, which can reduce the number of parameters that need to be updated. This sub-model achieves fine-tuning of specific scenarios by updating the weights of the adaptation layer.

[0066] The lighting demand extraction model adopts a hybrid structure of convolutional neural network and long short-term memory network. The convolutional neural network part uses 5 convolution layers, the convolution kernel size is 3×3, and the number of channels is 32, 64, 128, 256, and 512, respectively, to extract spatial features. The long short-term memory network part adopts a bidirectional structure, the hidden layer dimension is 256, and the time step is set to 24, corresponding to a time series of one day, to process time series features. The outputs of the two networks are fused through the attention mechanism to obtain the final feature vector.

[0067] The street light adaptive adjustment equation group module contains 5 core equations. The basic lighting equation takes into account the ambient light illumination and time factors, where the ambient light illumination threshold is set to 10,000 lux, the basic lighting coefficient ranges from 0.6 to 0.8, and the night compensation coefficient ranges from 0.2 to 0.4. The crowd response equation takes into account the crowd density and activity characteristics, where the maximum crowd density threshold is set to 5 people per square meter, and the weight coefficients are 0.2, 0.1, and 0.15 respectively. The vehicle response equation takes into account the vehicle density and speed factors, where the maximum vehicle density threshold is set to 200 vehicles per kilometer, and the reference speed is set to 60 kilometers per hour. The weather compensation equation takes into account visibility, precipitation and wind speed factors, where the standard visibility is set to 1,000 meters, and the weight coefficients are 0.15, 0.1, and 0.05 respectively. The energy efficiency optimization equation achieves a balanced optimization of multiple objectives.

[0068] The training process of the overall model of smart street lamp adaptive control is divided into four stages. In the pre-training stage, a large-scale data set containing 1 million samples is used, and the masked language model task is used for self-supervised learning. The batch size is set to 256, the initial value of the learning rate is 0.0001, and the cosine annealing strategy is used. In the fine-tuning stage, 100,000 annotated data of specific scenarios are used for training using supervised learning. In the validation and optimization stage, 10,000 independent validation data are used to optimize the model hyperparameters through the grid search method. In the deployment test stage, a one-month test is carried out in the actual environment to collect feedback data for model improvement.

[0069] The training process of the smart street lamp adaptive control sub-model includes two main steps: data collection and model training. In the data collection phase, real-time data for 3 months is collected within the coverage area of ​​the edge gateway, including environmental data, control effect feedback, etc. The data sampling frequency is 1 minute, and a local training set containing 50,000 samples is constructed. In the model training phase, the pre-trained total model weights are first loaded, and then the backbone network parameters are frozen, and only the low-rank adaptation layer parameters are updated. The training adopts the small batch stochastic gradient descent algorithm, with the batch size set to 32 and the learning rate set to 0.00001.

[0070] The street light control selection module selects different control strategies according to the system status. When the network delay is less than 100 milliseconds and the server computing load is less than 80%, the overall model and sub-model collaboration are adopted. When the network delay exceeds 100 milliseconds or the bandwidth is less than 1 megabit per second, switch to the equation group module control. When emergencies such as traffic accidents are detected, the local control mode is used first.

[0071] The core algorithms of the system include feature extraction, model training and control strategy generation. The feature extraction algorithm uses a combination of wavelet transform and Fourier transform to perform time-frequency domain analysis on environmental data and street light data. The wavelet basis function selects Durbin scale wavelet, and the number of decomposition layers is set to 4 layers. The Fourier transform uses the fast Fourier transform algorithm, and the window size is set to 1024 points. The model training algorithm uses the cross entropy loss function and adaptive moment estimation optimizer, and the learning rate uses the cosine annealing strategy, with an initial value of 0.001 and a minimum value of 0.00001. The control strategy generation algorithm uses a dynamic programming method to transform the control problem into an optimal path search problem.

[0072] Embodiment 5, The system's workflow is as follows: First, the edge gateway collects environmental data and street light status data through the sensor network. Environmental data includes pedestrian flow, vehicle flow, ambient light intensity, weather conditions, time period, etc., with a sampling frequency of 1 minute. Street light status data includes voltage, current, power, operating time, fault status, etc., with a sampling frequency of 1 second.

[0073] Next, the lighting demand extraction model extracts features from the collected data to obtain a feature vector. If the network conditions are met, the feature vector will be sent to the smart street light adaptive control overall model on the server side for processing to obtain the first street light adjustment vector.

[0074] Then, the smart street light adaptive control sub-model of the edge gateway fine-tunes the vector to obtain the second adjustment vector. Finally, the system generates specific control instructions based on the control vector and adjusts the street light through pulse width modulation signals and digital switching quantities.

[0075] The key parameters of the system are set as follows: The network communication uses the Advanced Message Queuing Protocol, and the service quality level is set to 1 to ensure reliable message transmission. Data compression uses a lossless compression algorithm with a compression rate of no less than 50%. Model quantization uses an 8-bit fixed-point solution, and the model size is compressed from the original 500 megabytes to 50 megabytes. The pulse width modulation frequency of the control signal is set to 20 kHz, and the duty cycle adjustment range is 10% to 90%.

[0076] The system's exception handling mechanisms include: When sensor data anomalies are detected, the Kalman filter algorithm is used to correct the data; when network communication interruption is detected, it automatically switches to local control mode; when street lamp failure is detected, the brightness of surrounding street lamps is adjusted through the load balancing algorithm to compensate. The process noise covariance matrix and measurement noise covariance matrix of the Kalman filter are calibrated through experimental data. The load balancing algorithm is implemented based on the minimum spanning tree method of graph theory.

[0077] The system's security mechanisms include: data transmission is encrypted using the transport layer security protocol with a key length of 256 bits; control instructions are authenticated using digital signature technology, using the elliptic curve digital signature algorithm; system access control adopts a role-based access control model that defines roles such as administrators, operators, and maintenance personnel.

[0078] The technical effects are mainly reflected in: through the combination of edge computing and artificial intelligence technology, precise control of smart street lamps is achieved; the layered architecture is adopted to reduce the system's computing load and network bandwidth requirements; through the coordination of multiple control strategies, the system's reliability and adaptability are improved; adaptive adjustment based on real-time data optimizes lighting effects and energy efficiency.

[0079] Specifically, the principle of the present invention is: 1. Design principles of hierarchical control architecture: The system adopts a layered architecture of "cloud-edge collaboration", which is designed based on the following considerations: The overall model is deployed in the cloud, which can fully utilize the computing resources of the data center to achieve global optimization; The sub-model is deployed at the edge, and lightweight adaptation is achieved through LoRA technology to reduce the communication burden; The empirical equations serve as the basis for local control and ensure the reliability and real-time performance of the system.

[0080] This hierarchical design complies with the "layered and graded" principle of the control system, which not only ensures the real-time performance of the system but also achieves global optimization.

[0081] 2. Principle of deep learning model selection: The overall model uses the Transformer structure based on its excellent feature extraction and long-range dependency modeling capabilities; The sub-model adopts the LoRA structure, which significantly reduces the number of parameters through low-rank decomposition and is suitable for edge deployment; The feature extraction model uses the CNN+LSTM structure to effectively process spatial and temporal features.

[0082] Model selection fully considers the needs of actual application scenarios and strikes a balance between performance and efficiency.

[0083] 3. Principle of adaptive control: The system's adaptive capabilities come from three levels: The online learning capability of deep learning models can continuously adapt to environmental changes; The fast fine-tuning mechanism of the LoRA structure enables fine adaptation to local scenes; The parameters of the empirical equation group are adaptive to provide basic control guarantee.

[0084] The multi-level adaptive mechanism ensures the stability of the system in various scenarios.

[0085] 4. Design principle of empirical equation group: The design of the system of equations is based on the following scientific principles: Principles of illumination physics: Consider the basic laws of light intensity attenuation and scattering; Principles of visual psychology: design compensation mechanisms based on the adaptation characteristics of the human eye; Traffic flow theory: mathematical description of the relationship between traffic density and speed; Control system theory: basic requirements such as stability and robustness.

[0086] The design of each equation has sufficient theoretical basis, which ensures the reliability of the control effect.

[0087] 5. Principle of fault tolerance mechanism: The reliability of the system is based on multiple protections: Model degradation strategy: automatically switch to a lightweight model when computing resources are insufficient; Communication backup mechanism: local control is enabled when network anomalies occur; Parameter constraints: Ensure that all control outputs are within a safe range.

[0088] Multiple protection mechanisms ensure the safe and reliable operation of the system.

[0089] Embodiment 6, System hardware architecture design, As the core control unit of the system, the edge gateway adopts a main controller based on an ARM Cortex-A53 quad-core processor with a main frequency of 1.5GHz, equipped with 2GB DDR4 running memory and 32GB eMMC storage, which can be expanded through a 64GB SD card and runs the Linux 4.19 LTS operating system. The gateway has complete communication capabilities, including a 4GLTE module supporting Band 1 / 3 / 5 / 8, a Wi-Fi module supporting IEEE 802.11 b / g / n / ac, a 10 / 100 / 1000 Mbps Ethernet interface, a LoRa module working in the 433MHz frequency band (communication distance up to 3km) and an RS485 interface for street light controller communication; it is also equipped with a variety of sensor interfaces, including 16-channel 16-bit ADC analog inputs, 32-channel optocoupler isolation digital inputs, 16-channel relay digital outputs and 8-channel high-speed counting pulse inputs, which can meet the data collection needs in complex scenarios.

[0090] The street light controller uses the STM32F407 microcontroller as the core, with a main frequency of 168MHz, equipped with 1MB Flash and 192KBRAM, and runs the FreeRTOS real-time operating system to ensure the real-time and reliability of control instructions; the power management module supports AC220V input and outputs DC 12V / 24V / 48V multiple voltages, with a conversion efficiency of not less than 92% and a power factor greater than 0.95; the dimming drive module uses a 2kHz PWM signal, has a 12-bit resolution, a maximum load power of 400W, and can achieve precise dimming control within the range of 0-100%; the sensor module integrates a light intensity sensor with a measurement range of 0-150000lux, a temperature sensor with a measurement range of -40℃~+85℃, a voltage and current detection unit with an accuracy of 0.5%, and a power factor measurement unit with an accuracy of 0.1%.

[0091] Communication protocol architecture design, A secure and reliable communication architecture is adopted between the server and the edge gateway. The application layer uses the MQTT / HTTP protocol, the transport layer adopts TLS 1.3 encryption, supports IPv4 / IPv6 network layer protocols, and can be flexibly deployed through various link layer access methods such as Ethernet, 4G or Wi-Fi; the MQTT topic design adopts a hierarchical structure, including control instruction topics, status report topics, alarm information topics and parameter configuration topics. All data is encapsulated in JSON format, including ISO8601 format timestamp, gateway unique identifier, data type and specific load data.

[0092] The communication between the edge gateway and the street light controller adopts a dual backup mechanism, and mainly communicates through the RS485 interface using the Modbus RTU protocol. The communication parameters are configured as 9600bps baud rate, 8 data bits, 1 stop bit, and no parity bit; at the same time, LoRa communication is equipped as a backup channel, working in the 433MHz frequency band, using 125kHz bandwidth, supporting dynamic adjustment of the spread factor of 7-12, coding rate 4 / 5, and transmission power 20dBm. The communication parameters can be adaptively adjusted according to the actual communication quality to ensure the reliability of data transmission.

[0093] Mathematical model design and implementation, 1. Basic lighting control model, The basic lighting control adopts a nonlinear dynamic model, and the core equation is: ,in Represents the basic brightness value (dimensionless, range [0,1]), is the ambient light intensity (in lux), is the maximum reference illuminance (10000lux), It is the 24-hour time (unit: h). is the basic lighting coefficient (value 0.7), is the night compensation coefficient (value is 0.3), is the time attenuation coefficient (value is 0.1), To obey The model realizes intelligent dimming through ambient light compensation and diurnal cycle changes, and its parameters are fitted and optimized based on historical data using nonlinear least squares method.

[0094] 2. Crowd flow response control model, The crowd flow response control model adopts an adaptive dynamic compensation strategy, and the core equation is: ,in Indicates the supplementary value of pedestrian lighting (range [0,0.3]), is the current crowd density (people / m²), is the maximum crowd density (5 people / m²), is the activity coefficient of the crowd (range [0,1]), is the maximum activity coefficient (value 1), is the weight coefficient (the values ​​are 0.2, 0.1, 0.15 respectively), is the change rate of human flow density, To obey The random error term of the distribution; the crowd density is calculated in real time through the video analysis algorithm: ,in Pixel The number detection function at is the detection area (m²), is the number of sampling points.

[0095] Traffic flow response control model, 3. Traffic flow response control adopts a multi-factor comprehensive evaluation model, and the core equation is: ,in Indicates the traffic lighting supplement value (range [0,0.4]), is the current traffic density (vehicles / km), is the maximum traffic density (200 vehicles / km), is the average vehicle speed (km / h), is the reference speed (60km / h), are weight coefficients (respectively 0.3 and 0.2), is the weight of the i-th car model (0.8 for small cars, 1.0 for medium cars, and 1.2 for large cars), is the distance attenuation coefficient (value range 0.05-0.15), is the distance from the i-th vehicle to the street light (m), To obey The vehicle density is monitored in real time by a geomagnetic sensor array, and the vehicle type recognition and speed measurement are performed in combination with a computer vision system to achieve accurate traffic flow feature extraction.

[0096] 4. Weather compensation control model, Weather compensation control adopts complex domain modeling method, and the core equation is: ,in Indicates the weather compensation value (range [0,0.2]), is the current visibility (m), is the standard visibility (1000m), is the precipitation intensity (mm / h), is the wind speed (m / s), is the relative humidity (range 0-1), is the weight coefficient (the values ​​are 0.15, 0.1, 0.05 respectively), is the humidity influence coefficient (value is 0.5), is an imaginary unit (used to represent the phase effect of wind speed), To obey The system collects environmental data in real time through a multi-parameter meteorological station, including comprehensive data from visibility sensors, rain gauges, anemometers and humidity sensors, to achieve intelligent compensation control for severe weather.

[0097] 5. Energy efficiency optimization control model, Energy efficiency optimization control adopts a multi-objective constraint optimization method, and the core equation is: ,in represents the final adjustment value (range [0,1]), is the current power (W), is the rated power (W), is the weight coefficient (values ​​are 1.0, 0.1 respectively), is the kth energy efficiency factor weight (value range 0.05-0.15), is the deviation value of the kth energy efficiency indicator, is the number of energy efficiency indicators, To obey The system monitors the operating status of street lamps in real time through the power parameter measurement module, and combines multiple energy efficiency evaluation indicators to achieve a dynamic balance between lighting effect and energy consumption.

[0098] Deep learning model architecture and system implementation plan, System hardware architecture, The system hardware architecture adopts a three-layer structure design, including the perception layer, network layer and computing layer. Each layer is equipped with corresponding hardware modules to ensure the stable operation and efficient control of the system: the perception layer uses a high-precision sensor array to realize multi-dimensional data collection, including XW-MS850 ambient light sensor (accuracy ±2lux, range 0-200000lux), HK-900T infrared array human detector (detection range 12m×12m, angular resolution 0.1°), TC-HDWF vehicle flow image collector (1920×1080 resolution, frame rate 30fps), WS-GP1 weather station (temperature accuracy ±0.1℃, humidity accuracy ±2%RH); the network layer uses industrial-grade Ethernet switches and 4G / 5G dual-mode communication modules, supports IEEE 802.3af / at standard PoE power supply, and the maximum power supply power of a single port is 30W; the computing layer uses edge computing servers (Intel Core i7-10700 processor, 32GB RAM, 512GB SSD) and NVIDIA Jetson AGX Xavier edge AI computing module (with 32TOPS AI performance and support for TensorRT acceleration) enables real-time reasoning and adaptive adjustment of the model.

[0099] Communication protocol architecture, The communication protocol architecture adopts a layered design to achieve seamless connection from the underlying hardware to the upper-layer applications: the physical layer adopts industrial Ethernet and 4G / 5G wireless networks, and supports IEEE 802.3 and 3GPP standards; the data link layer implements an efficient transmission mechanism based on CSMA / CD, and uses CRC32 check to ensure data integrity; the network layer uses the IPv6 protocol, supports QoS service quality assurance, and provides priority transmission for key control instructions; the transport layer adopts an improved TCP protocol, introduces a sliding window mechanism and congestion control algorithm to ensure the reliable transmission of large-scale data; the application layer adopts a publish-subscribe model based on MQTT to achieve flexible communication and data sharing between devices. The message body is encoded with Protocol Buffers, which has the characteristics of low latency and high compression rate. It also supports TLS 1.3 security encryption to protect data transmission security.

[0100] Detailed design of deep learning models, Overall model structure of smart street light adaptive control, The overall model of smart street light adaptive control adopts an improved Transformer structure, which includes the following key components: the multimodal feature encoder adopts a 6-layer stacking structure, each layer contains an 8-head self-attention mechanism (head dimension 64), and the position encoding adopts a sine-cosine function. , ,in is the position index, is the dimension index, The model dimension is 512; the multi-head attention calculation uses the scaled dot product attention mechanism ,in , , denote query, key, and value matrices respectively, The attention dimension is 64; the feature fusion adopts the residual connection structure ,in It is a two-layer feedforward network with a middle layer dimension of 2048 and GeLU as the activation function. The decoder also uses a 6-layer structure, each layer contains masked self-attention, encoder-decoder cross attention and feedforward network, and the output layer uses linear projection mapping to the control vector space.

[0101] Smart street light adaptive control sub-model structure, The sub-model is based on the LoRA structure to achieve efficient fine-tuning of the weight matrix of the original Transformer Perform low-rank decomposition: ,in is the pre-trained weight, and is the low-rank adaptation matrix, The rank parameter is 16; the model adopts a 3-layer lightweight structure, the hidden layer dimension is 256, the number of attention heads is 4, and each layer contains a feature selection module based on the gated attention mechanism ,in and is the learnable parameter matrix, is the sigmoid function, represents the element-by-element product; the output layer adopts a mixed expert system (Mixture of Experts) structure, which includes 4 expert networks. Each expert is responsible for generating control strategies in different scenarios. Dynamically distribute input to individual experts.

[0102] Lighting requirements extraction model structure, The lighting demand extraction model adopts a CNN+LSTM hybrid architecture. The CNN part uses ResNet-18 as the backbone network and contains 5 residual blocks. The calculation expression of each residual block is ,in and is the convolutional layer weight; the LSTM part adopts a bidirectional structure, and the hidden layer state update equation is: , , , , ,in , , They are forget gate, input gate and output gate respectively. is the unit state, is the hidden state, and is the weight matrix, is the bias term; the features of the two branches are integrated through the attention mechanism Perform dynamic weighted fusion, where and are learnable parameters.

[0103] Training data acquisition and preprocessing, The training data is acquired using a multi-source heterogeneous data acquisition solution: environmental data is collected in real time through a sensor network, including light intensity (sampling frequency 1Hz), temperature and humidity (sampling frequency 0.1Hz), and visibility (sampling frequency 0.1Hz); pedestrian flow data is acquired through infrared sensors and computer vision systems, and YOLOv5 is used for real-time target detection with an accuracy of more than 95%; vehicle flow data is acquired through geomagnetic detectors and video analysis systems, and the DeepSORT algorithm is used for multi-target tracking; street lamp operation data includes parameters such as voltage (accuracy ±0.1V), current (accuracy ±0.01A), and power factor (accuracy ±0.001); data preprocessing steps include: outlier detection using the DBSCAN clustering algorithm, adjacent data interpolation using cubic spline interpolation, time series alignment using the dynamic time warping (DTW) algorithm, and feature standardization using zero mean unit variance transformation .

[0104] Model training strategy and parameter configuration, The model adopts a multi-stage training strategy: the Adam optimizer is used in the pre-training stage, and the initial learning rate is , cosine annealing scheduling is used, batch size is 128, and training is 500 rounds; AdaBelief optimizer is used in the fine-tuning stage, and the learning rate is , using a progressive learning strategy, adding one type of scene data every 50 rounds; 5-fold cross validation is used in the verification phase, and the evaluation indicators include MAE, RMSE and MAPE; the loss function design adopts a multi-task learning framework: ,in To control the loss of precision, Optimize losses for energy consumption, The weight coefficient is determined by grid search to prevent overfitting, and weight decay is used and dropout ratio 0.1; the model quantization adopts 8-bit fixed-point scheme to adapt to edge device deployment.

[0105] System deployment and operation process, The system deployment adopts a containerized solution and uses Docker to build a microservice architecture, which includes modules such as data collection service, model inference service, and control execution service. The real-time control process is as follows: first, multi-source data is collected through the perception layer, and input into the lighting demand extraction model after preprocessing; the feature vector is collaboratively processed by the server-side total model and the edge-side sub-model to generate a preliminary control strategy; then the street light control selection module selects the execution path according to the network status and computing resources; finally, precise control is achieved through the PWM dimming module, and the control delay is less than 100ms; the system also includes a self-diagnosis mechanism to ensure stable operation through health status monitoring and fault warning; data storage uses the time series database InfluxDB, which supports high-concurrency writing and fast query; visual monitoring uses the Grafana platform to provide real-time monitoring and historical data analysis functions.

[0106] Equation system realization and optimization, Basic lighting equation fitting and optimization, The basic lighting equation uses nonlinear least squares method for parameter fitting, and the optimization objective function is: ,in is the actual measured value, is the predicted value, is the parameter set to be optimized; grid search is used to determine the optimal starting point for parameter initialization, and parameter constraints are set: , , ; Use the Levenberg-Marquardt algorithm to solve the optimal parameters, and the iteration termination condition is that the parameter change is less than Or the maximum number of iterations is 1000; the fitting accuracy is evaluated using the determination coefficient as a qualifying standard.

[0107] Fitting and optimization of crowd flow response equation, The parameter optimization of the crowd flow response equation adopts an adaptive parameter estimation method based on Bayesian optimization, and the objective function is designed as a multi-objective optimization problem that minimizes prediction error and energy consumption: ,in Supplementary values ​​for actual lighting, is the predicted value, is the power curve; Gaussian process regression is used to construct the probability model of parameter space, and the kernel function selects Matérn 5 / 2 kernel: ,in is the sample distance, is the length scale parameter; the parameter search space is set to: , , ; The optimization process uses the expected improvement criterion to select sampling points, and the iteration termination condition is that the improvement is less than 10 times in a row. or the maximum number of iterations 200 is reached.

[0108] Traffic flow response equation fitting and optimization, The traffic flow response equation adopts a piecewise adaptive parameter estimation strategy to divide the traffic density into low density ( vehicles / km), medium density ( vehicles / km) and high density ( Vehicles / km) in three intervals, and each interval has independent optimization parameters: ,in represents the density interval index, Represents the vehicle model index; the optimization objective function introduces a smooth transition constraint: ,in is the smoothing coefficient, which takes a value of 0.1; the parameter estimation is solved using the alternating direction multiplier method (ADMM): , , ,in is the augmented Lagrangian function, is the penalty parameter, and its initial value is set to 1.0.

[0109] Weather compensation equation fitting and optimization, The weather compensation equation parameter optimization adopts an integrated learning method based on particle swarm optimization (PSO), and the particle update equation is: , ,in is the inertia weight, using a linear decreasing strategy , , ; Learning factor , is a random number; the fitness function is designed as: ,in is the mean square error, To control the curve smoothness measure, is the energy consumption index, and the weight coefficient , , ; Parameter optimization range: , , , .

[0110] Energy efficiency optimization equation fitting and optimization, The energy efficiency optimization equation adopts the model predictive control (MPC) framework to solve the optimization problem in the sliding time domain: ,in To predict the time domain length 60, is the desired lighting curve, and is the weight matrix; system dynamic constraints: , , where the state variables include lighting intensity, power, and temperature; the control constraints are: , , ; The parameter identification adopts the subspace identification method, and the system order is determined by the AIC criterion; the solver adopts the OSQP algorithm, and the maximum number of iterations is set to 1000, and the solution accuracy is .

[0111] Embodiment 7, This embodiment takes a smart street lamp renovation project on a certain road section in a main urban area of ​​a city as an example to illustrate the specific implementation process of the present invention in detail.

[0112] The project covers a 12-kilometer section of the road, including 400 LED smart street lights, distributed in commercial areas, residential areas and school areas. The project took 6 months to complete system deployment and debugging.

[0113] Hardware configuration, This project adopts a regional deployment solution, and the hardware configuration is as follows.

[0114] 1. Street light configuration: Parameters Parameter Value illustrate Lamp Type LED Street Lights 200W rated power Light pole height 12 meters Double arm street light Lamp spacing 30 meters Standard Configuration Dimming range 20% to 100% PWM dimming Communication interface RS485 / 4G Dual mode communication 2. Edge gateway configuration: Configuration items Parameter Value illustrate CPU ARM Cortex A72 Quad-core 2.0GHz Memory 8GB LPDDR4 Industrial Grade storage 128GB SSD Industrial Grade Network Interface Gigabit Ethernet / 4G Dual-mode redundancy Power supply mode DC 24V Industrial Power Supplies 3. Sensor configuration: Sensor Type quantity Accuracy index Light sensor 80 ±2% FS Crowd Detector 40 95% accuracy Traffic Detector 40 98% accuracy Weather Station 8 Professional Electrical parameter acquisition 400 groups Level 0.5 Software deployment, The system software deployment adopts a layered architecture, and the specific configuration is described as follows.

[0115] 1. Server deployment: The system uses a server cluster with 8 NVIDIA A100 graphics cards as the computing platform to run the smart street light adaptive control model. The server-side model uses 4-bit quantization technology, and the single inference delay is controlled within 50 milliseconds. The system uses Docker container deployment to ensure the stability and scalability of the service. To ensure system reliability, a master-slave dual-machine hot backup mechanism is adopted, and the fault switching time does not exceed 3 seconds.

[0116] 2. Edge deployment: 40 edge gateways are evenly deployed in the project area, and each gateway is responsible for managing 10 street lamps. The edge runs the quantized smart street lamp adaptive control sub-model, the model size is controlled within 100MB, and the single inference time does not exceed 20 milliseconds. The system is pre-installed with the Linux operating system and adopts a real-time priority scheduling strategy.

[0117] 3. Data collection configuration: Data Types Sampling period Reporting cycle Light intensity 1 second 10 seconds Traffic data 1 second 30 seconds Traffic data 1 second 30 seconds Weather data 10 seconds 1 minute Electrical parameters 1 second 1 minute Control strategy implementation, This project formulates differentiated control strategies based on the characteristics of different regions. The specific implementation plan is described as follows.

[0118] 1. Commercial section control strategy: The commercial section is 3 kilometers long and covers large shopping malls and night markets. This section uses a relatively high basic lighting coefficient k1=0.8 and a night compensation coefficient k2=0.4. The crowd response weight β1 is increased to 0.25 to adapt to dense crowd scenes. The system automatically adjusts to high response mode on weekends and holidays, and the control delay is reduced to less than 30 milliseconds. According to measured data, the average daily power saving rate in this section reaches 28.5%.

[0119] 2. Residential area control strategy: The residential section is 6 kilometers long and includes multiple residential communities. This section uses a lower basic lighting coefficient k1=0.6, but increases the traffic response weight γ1 to 0.35. The system enters energy-saving mode after 23:00 at night, and the basic brightness is reduced to 30%, but the ability to respond quickly to moving targets is maintained. This section achieves an average daily power saving rate of 32.6%.

[0120] 3. School section control strategy: The school section is 3 kilometers long and includes two universities and three middle schools. The section adopts the pre-enhancement mode during the peak hours of school commuting, increasing the basic lighting intensity 5 minutes in advance. A new time period weight term is introduced into the passenger flow response equation, and the weight is increased by 50% during the school commuting period. The average daily power saving rate of this section reaches 30.8%.

[0121] like Figure 2 As shown, four key control parameters of commercial areas, residential areas and school areas are displayed, such as basic lighting coefficient k1, night compensation coefficient k2, pedestrian flow response weight β1, and vehicle flow response weight γ1.

[0122] System parameter configuration, According to the actual operation results, the key parameters of the system are configured as follows.

[0123] 1. Model training parameters: Parameters Overall model Sub-model Batch size 256 64 Learning Rate 0.001 0.0005 Training rounds 100 20 Quantization bit number 4-digit 8-bit 2. Control parameter configuration: Parameter Category Parameter Value Applicable scenarios Basic lighting factor 0.6 to 0.8 All time Night compensation factor 0.2 to 0.4 After 22:00 Traffic weight 0.2 to 0.25 Peak Traffic weight 0.3 to 0.35 All time Weather compensation factor 0.15 to 0.2 Inclement Weather Operation effect analysis, Since the system was put into trial operation in July 2023, a large amount of operation data has been collected. The data of a typical working day in December 2023 is now selected for analysis.

[0124] 1. Energy consumption comparison: Time Original system power / kW New system power / kW Energy saving rate / % Morning rush hour 80 62 22.5 Daytime 0 0 0 evening 80 56 30.0 Evening rush hour 80 68 15.0 late at night 80 42 47.5 2. Control performance indicators: Indicator Measured value Target value Response Delay 35ms 50ms Dimming accuracy ±2% ±5% System availability 99.95% 99.9% Local control switching time 100ms 200ms like Figure 3 As shown, the power variation curves of the original system and the new system within 24 hours are shown.

[0125] 3. Response to abnormal weather: The system performed well during the heavy snowfall on December 15, 2023. When visibility dropped to 200 meters, the system automatically activated the weather compensation mechanism and increased the lighting intensity by 40%, ensuring road safety. During this period, the system continued to operate stably without any failures.

[0126] like Figure 4 As shown, the changes in relative illumination intensity under different visibility conditions, such as sunny days, foggy days, rainy days, and snowy days, are demonstrated.

[0127] like Figure 5 As shown, various system performance indicators such as system availability, response delay, dimming accuracy, and control switching time are compared, including measured values, target values, peak values, and average values.

[0128] The above-mentioned embodiments are only preferred technical solutions of the present invention and should not be regarded as limitations of the present invention. The embodiments and features in the embodiments of the present application can be arbitrarily combined with each other without conflict. The protection scope of the present invention shall be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A smart street light adaptive control system based on edge gateway, characterized by: It includes a server and an edge gateway that are interconnected through a network; the server is provided with a smart street light adaptive control general model, and the edge gateway is provided with a smart street light adaptive control sub-model, a lighting demand extraction model, a street light control selection module and a street light adaptive adjustment equation group module; the lighting demand extraction model is used to extract features from environmental data and street light data to obtain feature vectors of each type of data, and after completing the feature vector extraction, all the obtained feature vectors are sent to the smart street light adaptive control general model in the server; the smart street light adaptive control general model calculates all the received feature vectors to obtain the first street light adjustment vector, and sends it to the edge gateway; the street light control selection module determines whether to send all the feature vectors to the smart street light adaptive control general model or to the street light adaptive adjustment equation group module; the street light adaptive adjustment equation group module calculates all the feature vectors obtained by the lighting demand extraction model through a preset empirical equation group.

2. According to the edge gateway-based smart street light adaptive control system of claim 1, it is characterized by: The smart street light adaptive control sub-model is a fine-tuned Lora model of the smart street light adaptive control overall model, which is used to enhance the adaptability of the adaptive control of street lights connected to the edge gateway; the smart street light adaptive control sub-model calculates the first street light adjustment vector to obtain the second street light adjustment vector for generating street light control instructions.

3. The edge gateway-based smart street light adaptive control system according to claim 1 is characterized in that: The overall model of the smart street lamp adaptive control adopts the Transformer encoder-decoder structure; the sub-model of the smart street lamp adaptive control adopts the lightweight LoRA structure; and the lighting demand extraction model adopts the CNN+LSTM hybrid structure.

4. The edge gateway-based smart street light adaptive control system according to claim 1 is characterized in that: The street light adaptive adjustment equation group includes basic lighting equation, pedestrian flow response equation, vehicle flow response equation, weather compensation equation and energy efficiency optimization equation; the street light adjustment vector includes brightness value, start and stop time, gradient rate and group linkage parameter.

5. The edge gateway-based smart street light adaptive control system according to claim 4 is characterized in that: The basic lighting equation is as follows: ； In the formula, is the basic brightness value, ranging from [0,1]; is the ambient light intensity (lux); is the maximum reference illuminance, which is 10000 lux; It is the 24-hour time (h); is the basic lighting coefficient, taking the value [0.6,0.8]; is the night compensation coefficient, with a value of [0.2, 0.4]; is the time attenuation coefficient, which takes a value of 0.1; is the error term.

6. The edge gateway-based smart street light adaptive control system according to claim 4 is characterized in that: The flow response equation is as follows: ； In the formula, The supplementary value of pedestrian lighting, range [0,0.3]; is the current crowd density (people / m²); is the maximum crowd density, which is 5 people / m²; is the activity coefficient of the population; is the maximum activity coefficient, which takes the value of 1; is the weight coefficient, which takes values ​​of 0.2, 0.1, and 0.15 respectively; is the change rate of human flow density; is the error term.

7. The edge gateway-based smart street light adaptive control system according to claim 4 is characterized in that: The traffic flow response equation is as follows: ； In the formula, Supplementary value for traffic lighting, range [0,0.4]; is the current traffic density (vehicles / km); is the maximum traffic density, which is 200 vehicles / km; is the average vehicle speed (km / h); is the reference speed, which is 60km / h; are weight coefficients, taking values ​​of 0.3 and 0.2 respectively; is the weight of the i-th vehicle model; is the distance attenuation coefficient; is the distance from the i-th vehicle to the street light (m); is the number of vehicle types; is the error term.

8. The edge gateway-based smart street light adaptive control system according to claim 4 is characterized in that: The weather compensation equation is as follows: ； In the formula, is the weather compensation value, ranging from [0,0.2]; is the current visibility (m); is the standard visibility, with a value of 1000m; is the precipitation intensity (mm / h); is the wind speed (m / s); is the relative humidity; is the weight coefficient, which takes values ​​of 0.15, 0.1, and 0.05 respectively; is the humidity influence coefficient, the value is 0.5; is an imaginary unit used to represent the phase effect of wind speed; is the error term.

9. The edge gateway-based smart street light adaptive control system according to claim 4 is characterized in that: The energy efficiency optimization equation is as follows: ； In the formula, is the final adjustment value, ranging from [0,1]; is the current power (W); is the rated power (W); are weight coefficients, taking values ​​of 1.0 and 0.1 respectively; is the kth energy efficiency factor weight; is the deviation value of the kth energy efficiency indicator; is the number of energy efficiency indicators; is the error term.

10. The edge gateway-based smart street light adaptive control system according to any one of claims 1 to 9, characterized in that: The selection criteria of the street light control selection module are: when the network communication is normal and the computing resources are sufficient, the smart street light adaptive control general model + smart street light adaptive control sub-model collaborative approach is used; when the network delay exceeds the threshold or the bandwidth is limited, the street light adaptive adjustment equation group module is used; when an emergency is detected, the local control mode, that is, the street light adaptive adjustment equation group module, is used; at the same time, the edge gateway adjusts the street light through PWM dimming signals and digital switching quantities.

Images