Road intelligent lighting control method integrating AI edge control

Abstract

The invention discloses a road intelligent lighting control method integrating AI edge control, and relates to the technical field of lighting control. The energy-saving effect is improved, compared with a traditional PID control method, 38.7% of energy efficiency can be improved through the scheme, and invalid energy consumption is effectively avoided through intelligent illumination control; according to the method, the illumination quality is improved, the intelligent control strategy is combined with the traffic flow, the environment change and the spatio-temporal data analysis, the illumination intensity can be optimized on the premise of ensuring the road safety, the illumination standard reaching rate is improved by 25.3%, and the road illumination is more intelligent and accurate.

Description

Technical Field

[0001] The present invention relates to the field of lighting control technology, and specifically to an intelligent highway lighting control method integrating AI edge control. Background Art

[0002] With the acceleration of urbanization, the intelligentization of highway lighting has gradually become one of the key technologies to improve road safety, energy conservation and environmental protection.

[0003] Traditional highway lighting control systems rely on timed switches, light intensity, or traffic flow to adjust lighting. However, these methods often lack dynamic adaptation to multiple factors, such as real-time traffic conditions and ambient lighting changes, resulting in energy waste and inefficient control. Specific issues include the following:

[0004] Low energy efficiency and waste of resources:

[0005] Traditional highway lighting systems typically rely on fixed schedules or simple light sensors to control the on / off switching of streetlights. This fixed control strategy ignores factors such as actual traffic volume, weather variations, and daytime and nighttime light levels. As a result, high lighting intensity is maintained during periods of low traffic or good weather, resulting in significant energy waste.

[0006] Lack of real-time dynamic adaptation capabilities:

[0007] Existing lighting systems often lack the flexibility to respond to real-time traffic conditions and environmental changes. For example, traditional systems often fail to intelligently adjust lighting brightness based on factors such as road speed, traffic volume, and weather changes. This results in insufficient lighting in accident-prone areas or during unusual weather conditions, while consuming excessive energy during periods of low traffic.

[0008] Data processing and feedback delay:

[0009] Many existing smart lighting systems rely on cloud computing platforms to process data from various sensors. However, this cloud-based data transmission model introduces high latency, making the system's real-time performance poor, especially when rapid response is required. Data must be transmitted to the cloud for calculation and feedback, increasing response time and impacting the lighting system's efficiency.

[0010] The control strategy is single and cannot be comprehensively optimized:

[0011] While some current intelligent lighting systems incorporate intelligent algorithms for lighting control, most still rely on simple preset rules or timing strategies, lacking comprehensive optimization solutions tailored to different environments, road conditions, and traffic conditions. Traditional control methods fail to fully account for the dynamic changes in multiple factors, such as traffic flow, road conditions, and weather, limiting the effectiveness of intelligent control.

[0012] Poor distributed control and coordination capabilities:

[0013] Many existing systems still rely on local or centralized control, lacking coordinated management of multiple streetlight control nodes. This can lead to localized adjustments and inconsistencies, a problem that becomes more pronounced in large-scale deployments, preventing global optimization.

[0014] Therefore, a new solution to the above problems needs to be proposed. Summary of the Invention

[0015] The purpose of the present invention is to provide an intelligent highway lighting control method integrating AI edge control to solve the technical problems raised in the background technology.

[0016] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent highway lighting control integrating AI edge control, comprising at least the following steps:

[0017] S1: First, build the cloud-edge-end three-level architecture;

[0018] S2: Design a spatiotemporal feature extractor within the cloud-edge-device three-level architecture. This spatiotemporal feature extractor uses an improved graph convolutional gated recurrent unit to extract spatial features through graph convolution operations. This extractor also incorporates dynamic changes in time series to effectively capture traffic flow propagation characteristics.

[0019] S3: Build a dynamic lighting demand prediction model that combines multimodal data on traffic flow, ambient light, and weather, and uses an attention mechanism to capture the dynamic coupling relationship between different factors.

[0020] S4: Build a distributed model predictive controller (DMPC). The objective function of the DMPC is to minimize the error and the change of the control input to ensure the execution effect of the control strategy at multiple moments. The constraints ensure the feasibility and practicality of the distributed model predictive controller.

[0021] S5: To improve the inference efficiency of edge computing nodes, edge computing optimization is designed. The optimization strategies adopted by the edge computing optimization are lightweight model compression and mixed precision quantization.

[0022] S6: In order to optimize the energy efficiency of lighting, the nonlinear characteristics of LED lamps are considered and an energy efficiency optimization model is proposed;

[0023] S7: Build a safety enhancement mechanism in the cloud-edge-end three-level architecture to ensure the stability of lighting during driving and the lighting response in emergency situations. The safety enhancement mechanism includes lighting gradient constraints and emergency lighting trigger conditions.

[0024] Furthermore, the cloud-edge-end three-level architecture in S1 includes the end side, edge side and cloud side.

[0025] The device integrates millimeter-wave radar, low-light cameras, and photosensors into smart light poles, which monitor traffic flow, light intensity, and environmental changes in real time. The data collected by these sensors is fed into a dynamic lighting demand prediction model, where it is used to support intelligent lighting control decisions.

[0026] The edge side deploys edge computing nodes on every kilometer of road. The nodes run lightweight spatiotemporal prediction models, can quickly process data from the end side, and make local decisions. The edge computing architecture reduces communication latency and improves system response speed.

[0027] The cloud side is responsible for global parameter optimization and model update. The cloud side implements global scheduling of the system, optimizes the lighting control strategy of each node, and regularly updates the model to the edge node, which can ensure that the system is optimized according to global needs and maintain the intelligence of the system.

[0028] Furthermore, the spatiotemporal feature extractor in S2 adopts an improved graph convolution gated recurrent unit to extract spatial features through graph convolution operations and combines the dynamic changes of time series;

[0029] The spatiotemporal feature update formula is as follows:

[0030] h t =GRU-Cell(GCONV(X t ,A),h t-1 )

[0031] The product kernel of the volume is defined as:

[0032] GCONV(X,A)=σ((D -0.5 AD -0.5 )XW g )

[0033] Where: A is the link adjacency matrix, which represents the connection relationship between different links; D is the degree matrix, which represents the connectivity of each link; W g is the learnable graph convolution weight; v is the activation function for nonlinear transformation; h t ∈R d is the hidden state vector at time t, and dimension d represents the feature encoding dimension;

[0034] Furthermore, the formula of the attention mechanism is as follows:

[0035]

[0036] Where: Q = W q [h t ; E] is the query vector, combined with the hidden state h at time t t Ambient lighting characteristics E; K = W k [S;W] is the key vector, combining the traffic speed matrix S and the weather code W; V = W v [h t ] is a value vector, representing the hidden state at the current moment.

[0037] Furthermore, in the distributed model predictive control, the goal is to minimize the system output error and the change of the control input by optimizing the control input, ensuring that the control strategy can be effectively executed at multiple moments, thereby forming an optimization objective function;

[0038] The optimization objective function includes the error term and the change term of the control input;

[0039] The error term is the difference between the expected reference trajectory r t+k The deviation between where y t+k is the actual output of the system, Q is the error weighting matrix;

[0040] The change term of the control input is the sum of the squares of the control input changes Used to suppress large fluctuations in control input, R is the input change weighting matrix;

[0041] Therefore, the optimization goal is:

[0042]

[0043] To ensure the feasibility and practicality of the distributed model predictive controller, the following constraints are introduced:

[0044] The value range of the control input u is limited to:

[0045] 0.3L max ≤u≤L max

[0046] Among them L max is the maximum limit of the control input;

[0047] The limits on control input changes are:

[0048] |Δu|≤0.2L max / s

[0049] Ensure that changes in control inputs are not too drastic;

[0050] The total control input does not exceed the grid power limit:

[0051] ∑u i ≤P grid (t)

[0052] That is, the total power of the control input cannot exceed the grid power at time t;

[0053] The alternating direction multiplier method is used as the optimization algorithm to decompose and optimize the control input;

[0054] The updating steps of the alternating direction multiplier method are as follows:

[0055] Control input update:

[0056]

[0057] Intermediate variable update:

[0058]

[0059] Lagrange multiplier update:

[0060]

[0061] Among them, L ρ (u i ,z k ,λ k ) is the Lagrangian function; C is the projection operator; is the Lagrange multiplier; ρ is the step size parameter; N is the number of nodes; u i represents the control input or decision variable; z k represents an intermediate variable or a co-variable; k is the Lagrange multiplier, which represents the influence of the constraints on the objective function in the optimization problem.

[0062] Furthermore, the lightweight model compression in S5 is to reduce storage and computing overhead by sparsely compressing the model weights;

[0063] The sparseness strategy is achieved through the Bernoulli distribution M ij ~Bernoulli(p) is used to randomly select weights so that most weights are close to zero, thereby reducing the storage and computation requirements of the model;

[0064] The sparsification formula of the sparsification strategy is:

[0065] W pruned =W⊙M,M ij~Bernoulli(p)

[0066] Where p = 1-exp(-α|W ij |) controls the sparsity; α is a hyperparameter that controls the degree of sparsity, W ij is the weight in the model; W is the original weight matrix; ⊙ represents element-by-element multiplication; M is a binary matrix where each element M ij , which is either 1 or 0, determines the original weight.

[0067] Furthermore, the mixed precision quantization in S5 converts floating-point numbers into fixed-point numbers with lower precision. The mixed precision quantization adjusts the precision by controlling the quantization step size Δ. The formula is:

[0068]

[0069] Where b is the bit width after quantization; max|x| is the maximum absolute value in the data; and Q(x) is the quantized value.

[0070] Furthermore, the energy efficiency optimization model takes into account the nonlinear characteristics of LED lamps. The optimization goal is to reduce the energy consumption of the lighting system while ensuring the lighting effect. The formula of the energy efficiency optimization model is as follows:

[0071]

[0072] Among them: a i 、b i and c i These three parameters represent the characteristics of the i-th luminaire; is the nonlinear part, which represents the relationship between energy consumption and the control input of the lamp setting, u i is the control input of the i-th luminaire; β is the smoothing coefficient, which is used to adjust the smoothness of the control input change. A larger β value helps reduce the volatility of the control input and prevent the impact of brightness fluctuations on the visual effect; Var(u) is the variance of the control input, which represents the fluctuation amplitude of the control input. In order to smooth the lighting changes, it reduces the rapid changes of the control input, improves the user experience and energy saving effect.

[0073] The goal of the energy efficiency optimization model is to optimize the energy efficiency of the lighting system by adjusting the control input u i To minimize total energy consumption while ensuring smooth lighting and avoiding energy waste or undesirable lighting effects due to unstable control input.

[0074] Furthermore, the illumination gradient constraint is:

[0075]

[0076] Where: L i and L j Indicates the lighting brightness at different locations; It is the lighting brightness gradient, which indicates the brightness difference between two adjacent lamps. Controlling the lighting gradient helps avoid sudden changes in lighting. E represents the connection between lamps, usually referring to adjacent lamps in a lamp layout.

[0077] The constraints are:

[0078] |L i -L j |≤0.3max(L i ,L j )

[0079] The brightness difference between adjacent lamps is required to not exceed 30% of the maximum brightness. This constraint ensures the smoothness of the lighting transition and avoids discomfort or interference to the driver caused by large brightness changes.

[0080] Furthermore, the triggering condition of the emergency lighting is:

[0081] L emergency =max(L normal ,L min +kv 2 )

[0082] Where: L emergency is the brightness of emergency lighting; L normal is the lighting brightness under normal conditions; L min is the minimum lighting brightness, ensuring it does not drop below a certain minimum brightness in an emergency. κ is a coefficient related to vehicle speed, representing the impact of speed on emergency lighting brightness. Faster vehicle speeds result in brighter emergency lighting, ensuring sufficient brightness for safety even at high speeds. v is vehicle speed.

[0083] Compared with the prior art, the present invention has the following beneficial effects:

[0084] 1. The present invention improves energy-saving effects. Compared with the traditional PID control method, the solution of the present invention can achieve a 38.7% improvement in energy efficiency. Through intelligent lighting control, it effectively avoids ineffective energy consumption.

[0085] 2. This invention improves lighting quality. By combining intelligent control strategies with traffic flow, environmental changes, and spatiotemporal data analysis, it can optimize lighting intensity while ensuring road safety, increasing the illumination compliance rate by 25.3%, making road lighting more intelligent and precise.

[0086] 3. The present invention provides real-time response and low latency. The deployment of edge computing nodes enables the entire system to react quickly after receiving data, reducing the control response delay to less than 120 milliseconds, ensuring timely response to changes in traffic conditions.

[0087] 4. The present invention is designed to ensure road safety. Through the combined effect of the emergency lighting trigger mechanism and the light gradient constraint, it ensures that in emergencies such as traffic accidents and bad weather, the lighting can be quickly adjusted to ensure the safety of the driver. DETAILED DESCRIPTION

[0088] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0089] A method for intelligent highway lighting control integrating AI edge control includes at least the following steps:

[0090] S1: First, build the cloud-edge-end three-level architecture;

[0091] S2: Design a spatiotemporal feature extractor within the cloud-edge-device three-level architecture. This spatiotemporal feature extractor uses an improved graph convolutional gated recurrent unit to extract spatial features through graph convolution operations. This extractor also incorporates dynamic changes in time series to effectively capture traffic flow propagation characteristics.

[0092] S3: Build a dynamic lighting demand prediction model. This model combines multimodal data such as traffic flow, ambient light, and weather, and uses an attention mechanism to capture the dynamic coupling relationship between different factors.

[0093] S4: Build a distributed model predictive controller (DMPC). The objective function in DMPC is to minimize the error and the change of the control input to ensure the execution effect of the control strategy at multiple times. The constraints ensure the feasibility and practicality of the distributed model predictive controller.

[0094] S5: To improve the inference efficiency of edge computing nodes, we design edge computing optimization. The optimization strategies used in edge computing optimization are lightweight model compression and mixed precision quantization.

[0095] S6: In order to optimize the energy efficiency of lighting, the nonlinear characteristics of LED lamps are considered and an energy efficiency optimization model is proposed;

[0096] S7: Build a safety enhancement mechanism in the cloud-edge-end three-level architecture to ensure the stability of lighting during driving and the lighting response in emergency situations. The safety enhancement mechanism includes lighting gradient constraints and emergency lighting trigger conditions.

[0097] The graph convolutional gated recurrent unit effectively captures the characteristics of traffic propagation and can perform accurate spatiotemporal feature extraction in complex road topology.

[0098] The attention mechanism improves the adaptability and accuracy of the model by combining the dynamic coupling of meteorological factors and traffic conditions.

[0099] Distributed MPC can ensure global optimization and local rapid response, and adapt to dynamically changing traffic conditions.

[0100] Physical property constraint design ensures that lighting control can save energy while ensuring driving safety.

[0101] Edge computing optimization greatly improves edge-side reasoning efficiency and reduces latency through lightweight compression and mixed-precision quantization technology.

[0102] The cloud-edge-end three-level architecture in S1 includes the end side, edge side, and cloud side.

[0103] Smart light poles, equipped with integrated millimeter-wave radar, low-light cameras, and photosensors, monitor traffic flow, light intensity, and environmental changes in real time. Data collected by these sensors is fed into a dynamic lighting demand prediction model, where it is used to support intelligent lighting control decisions.

[0104] On the edge side, edge computing nodes are deployed on every kilometer of road. These nodes run lightweight spatiotemporal prediction models, can quickly process data from the end side, and make local decisions. This edge computing architecture reduces communication latency and improves system response speed.

[0105] The cloud side is responsible for global parameter optimization and model update. It implements global scheduling of the system, optimizes the lighting control strategy of each node, and regularly updates the model to the edge node, ensuring that the system is optimized according to global needs and maintaining the intelligence of the system.

[0106] The spatiotemporal feature extractor in S2 uses an improved graph convolution gated recurrent unit to extract spatial features through graph convolution operations and incorporate the dynamic changes of time series;

[0107] The spatiotemporal feature update formula is as follows:

[0108] h t =GRU-Cell(GCONV(X t ,A),h t-1 )

[0109] The product kernel of the volume is defined as:

[0110] GCONV(X,A)=σ((D -0.5AD -0.5 )XW g )

[0111] Where: A is the link adjacency matrix, which represents the connection relationship between different links; D is the degree matrix, which represents the connectivity of each link; W g is the learnable graph convolution weight; σ is the activation function for nonlinear transformation; h t ∈R d is the hidden state vector at time t, and dimension d represents the feature encoding dimension;

[0112] The formula of the attention mechanism is as follows:

[0113]

[0114] Where: Q = W q [h t ; E] is the query vector, combined with the hidden state h at time t t Ambient lighting characteristics E; K = W k [S;W] is the key vector, combining the traffic speed matrix S and the weather code W; V = W v [h t ] is a value vector, representing the hidden state at the current moment.

[0115] In distributed model predictive control, the goal is to minimize the system output error and the change of the control input by optimizing the control input, ensuring that the control strategy can be effectively executed at multiple times, thereby forming an optimization objective function;

[0116] The optimization objective function includes the error term and the change term of the control input;

[0117] The error term is the difference between the expected reference trajectory r t+k The deviation between where y t+k is the actual output of the system, Q is the error weighting matrix;

[0118] The change term of the control input is the sum of the squares of the control input changes Used to suppress large fluctuations in control input, R is the input change weighting matrix;

[0119] Therefore, the optimization goal is:

[0120]

[0121] To ensure the feasibility and practicality of the distributed model predictive controller, the following constraints are introduced:

[0122] The value range of the control input u is limited to:

[0123] 0.3Lmax ≤R≤L max

[0124] Among them L max is the maximum limit of the control input;

[0125] The limits on control input changes are:

[0126] |Δu|≤0.2L max / s

[0127] Ensure that changes in control inputs are not too drastic;

[0128] The total control input does not exceed the grid power limit:

[0129] ∑u i ≤P grid (t)

[0130] That is, the total power of the control input cannot exceed the grid power at time t;

[0131] The alternating direction multiplier method is used as the optimization algorithm to decompose and optimize the control input;

[0132] The update steps of the alternating direction multiplier method are as follows:

[0133] Control input update:

[0134]

[0135] Intermediate variable update:

[0136] Lagrange multiplier update:

[0137]

[0138] Among them, L ρ (u i ,z k ,λ k ) is the Lagrangian function; Π C is the projection operator; is the Lagrange multiplier; ρ is the step size parameter; N is the number of nodes; u i represents the control input or decision variable; z k represents an intermediate variable or a co-variable; k is the Lagrange multiplier, which represents the influence of the constraints on the objective function in the optimization problem.

[0139] The lightweight model compression in S5 is to reduce storage and computing overhead by sparsely compressing the model weights;

[0140] The sparseness strategy is achieved through the Bernoulli distribution M ij~Bernoulli(p) is used to randomly select weights so that most weights are close to zero, thereby reducing the storage and computation requirements of the model;

[0141] The sparsification formula of the sparsification strategy is:

[0142] W pruned =W⊙M,M ij ~Bernoulli(p)

[0143] Where p = 1-exp(-α|W ij |) controls the sparsity; α is a hyperparameter that controls the degree of sparsity, W ij is the weight in the model; W is the original weight matrix; ⊙ represents element-by-element multiplication; M is a binary matrix where each element M ij , which is either 1 or 0, determines the original weight.

[0144] The mixed precision quantization in S5 converts floating-point numbers into fixed-point numbers with lower precision. Mixed precision quantization adjusts the precision by controlling the quantization step size Δ. The formula is:

[0145]

[0146] Where b is the bit width after quantization; max|x| is the maximum absolute value in the data; and Q(x) is the quantized value.

[0147] The energy efficiency optimization model takes into account the nonlinear characteristics of LED lamps. The optimization goal is to reduce the energy consumption of the lighting system while ensuring the lighting effect. The formula of the energy efficiency optimization model is as follows:

[0148]

[0149] Among them: a i 、b i and c i These three parameters represent the characteristics of the i-th luminaire; is the nonlinear part, which represents the relationship between energy consumption and the control input of the lamp setting, u i is the control input of the i-th luminaire; β is the smoothing coefficient, which is used to adjust the smoothness of the control input change. A larger β value helps reduce the volatility of the control input and prevent the impact of brightness fluctuations on the visual effect; Var(u) is the variance of the control input, which represents the fluctuation amplitude of the control input. In order to smooth the lighting changes, it reduces the rapid changes of the control input, improves the user experience and energy saving effect.

[0150] The goal of the energy efficiency optimization model is to optimize the energy efficiency of the lighting system by adjusting the control input u iTo minimize total energy consumption while ensuring smooth lighting and avoiding energy waste or undesirable lighting effects due to unstable control input.

[0151] The illumination gradient constraint is:

[0152]

[0153] Where: L i and L j Indicates the lighting brightness at different locations; It is the lighting brightness gradient, which indicates the brightness difference between two adjacent lamps. Controlling the lighting gradient helps avoid sudden changes in lighting. E represents the connection between lamps, usually referring to adjacent lamps in a lamp layout.

[0154] The constraints are:

[0155] |L i -L j |≤0.3max(L i ,L j )

[0156] The brightness difference between adjacent lamps is required to not exceed 30% of the maximum brightness. This constraint ensures the smoothness of the lighting transition and avoids discomfort or interference to the driver caused by large brightness changes.

[0157] The triggering conditions for emergency lighting are:

[0158] L emergency =max(L normal ,L min +kv 2 )

[0159] Where: L emergency is the brightness of emergency lighting; L normal is the lighting brightness under normal conditions; L min is the minimum lighting brightness, ensuring it does not drop below a certain minimum brightness in an emergency. κ is a coefficient related to vehicle speed, representing the impact of speed on emergency lighting brightness. Faster vehicle speeds result in brighter emergency lighting, ensuring sufficient brightness for safety even at high speeds. v is vehicle speed.

[0160] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations that come within the meaning and range of equivalents of the claims be embraced therein.

Claims

1. A method for intelligent highway lighting control integrating AI edge control, characterized by: At least the following steps are included: S1: First, build the cloud-edge-end three-level architecture; S2: Design a spatiotemporal feature extractor within the cloud-edge-device three-level architecture. This spatiotemporal feature extractor uses an improved graph convolutional gated recurrent unit to extract spatial features through graph convolution operations. This extractor also incorporates dynamic changes in time series to effectively capture traffic flow propagation characteristics. S3: Build a dynamic lighting demand prediction model that combines multimodal data on traffic flow, ambient light, and weather, and uses an attention mechanism to capture the dynamic coupling relationship between different factors. S4: Build a distributed model predictive controller (DMPC). The objective function of the DMPC is to minimize the error and the change of the control input to ensure the execution effect of the control strategy at multiple moments. The constraints ensure the feasibility and practicality of the distributed model predictive controller. S5: To improve the inference efficiency of edge computing nodes, edge computing optimization is designed. The optimization strategies adopted in the edge computing optimization are lightweight model compression and mixed precision quantization. S6: In order to optimize the energy efficiency of lighting, the nonlinear characteristics of LED lamps are considered and an energy efficiency optimization model is proposed; S7: Build a safety enhancement mechanism in the cloud-edge-end three-level architecture to ensure the stability of lighting during driving and the lighting response in emergency situations. The safety enhancement mechanism includes lighting gradient constraints and emergency lighting trigger conditions.

2. The method for intelligent highway lighting control integrating AI edge control according to claim 1, characterized in that: The cloud-edge-end three-level architecture in S1 includes the end side, edge side and cloud side The device integrates millimeter-wave radar, low-light cameras, and photosensors into smart light poles, which monitor traffic flow, light intensity, and environmental changes in real time. The data collected by these sensors is fed into a dynamic lighting demand prediction model, where it is used to support intelligent lighting control decisions. The edge side deploys edge computing nodes on every kilometer of road. The nodes run lightweight spatiotemporal prediction models, can quickly process data from the end side, and make local decisions. The edge computing architecture reduces communication latency and improves system response speed. The cloud side is responsible for global parameter optimization and model update. The cloud side implements global scheduling of the system, optimizes the lighting control strategy of each node, and regularly updates the model to the edge node, which can ensure that the system is optimized according to global needs and maintain the intelligence of the system.

3. The method for intelligent highway lighting control integrating AI edge control according to claim 1, characterized in that: The spatiotemporal feature extractor in S2 adopts an improved graph convolution gated recurrent unit to extract spatial features through graph convolution operations and combines the dynamic changes of time series; The spatiotemporal feature update formula is as follows: h t =GRU-Cell(GCONV(X t ,A),h t-1 ) The product kernel of the volume is defined as: GCONV(X,A)=σ((D -0.5 AD -0.5 )XW g ) Where: A is the link adjacency matrix, which represents the connection relationship between different links; D is the degree matrix, which represents the connectivity of each link; W g is the learnable graph convolution weight; σ is the activation function for nonlinear transformation; h t ∈R d is the hidden state vector at time t, and dimension d represents the feature encoding dimension.

4. The method for intelligent highway lighting control integrating AI edge control according to claim 3 is characterized by: The formula of the attention mechanism is as follows: Where: Q = W q [h t ; E] is the query vector, combined with the hidden state h at time t t Ambient lighting characteristics E; K = W k [S;W] is the key vector, combining the traffic speed matrix S and the weather code W; V = W v [h t ] is a value vector, representing the hidden state at the current moment.

5. The method for intelligent highway lighting control integrating AI edge control according to claim 1, characterized in that: In the distributed model predictive control, the goal is to minimize the system output error and the change of the control input by optimizing the control input, ensuring that the control strategy can be effectively executed at multiple times, thereby forming an optimization objective function; The optimization objective function includes the error term and the change term of the control input; The error term is the difference between the expected reference trajectory r t+k The deviation between where y t+k is the actual output of the system, Q is the error weighting matrix; The change term of the control input is the sum of the squares of the control input changes Used to suppress large fluctuations in control input, R is the input change weighting matrix; Therefore, the optimization goal is: To ensure the feasibility and practicality of the distributed model predictive controller, the following constraints are introduced: The value range of the control input u is limited to: <h2 style=";text-align:left;direction:ltr">0.3L<h2 style=";text-align:left;direction:ltr"> max <h2 style=";text-align:left;direction:ltr"> ≤u≤L<h2 style=";text-align:left;direction:ltr"> max Among them L max is the maximum limit of the control input; The limits on control input changes are: |Δu|≤0.2L max / s Ensure that changes in control inputs are not too drastic; The total control input does not exceed the grid power limit: ∑u i ≤P grid (t) That is, the total power of the control input cannot exceed the grid power at time t; The alternating direction multiplier method is used as the optimization algorithm to decompose and optimize the control input; The updating steps of the alternating direction multiplier method are as follows: Control input update: Intermediate variable update: Lagrange multiplier update: Among them, L ρ (u i ,z k ,λ k ) is the Lagrangian function; Π C is the projection operator; is the Lagrange multiplier; ρ is the step size parameter; N is the number of nodes; u i represents the control input or decision variable; z k represents an intermediate variable or a co-variable; k is the Lagrange multiplier, which represents the influence of the constraints on the objective function in the optimization problem.

6. The method for intelligent highway lighting control integrating AI edge control according to claim 1, characterized in that: The lightweight model compression in S5 is to reduce storage and computing overhead by sparsely compressing the model weights; The sparseness strategy is achieved through the Bernoulli distribution M ij ~Bernoulli(p) is used to randomly select weights so that most weights are close to zero, thereby reducing the storage and computation requirements of the model; The sparsification formula of the sparsification strategy is: W pruned =W⊙M,M ij ~Bernoulli(p) Where p = 1-exp(-α|W ij |) controls the sparsity; α is a hyperparameter that controls the degree of sparsity, W ij is the weight in the model; W is the original weight matrix; ⊙ represents element-by-element multiplication; M is a binary matrix, where each element M ij , which is either 1 or 0, determines the original weight.

7. The method for intelligent highway lighting control integrating AI edge control according to claim 6, characterized in that: The mixed precision quantization in S5 converts floating-point numbers into fixed-point numbers with lower precision. The mixed precision quantization adjusts the precision by controlling the quantization step size Δ. The formula is: Where b is the bit width after quantization; max|x| is the maximum absolute value in the data; and Q(x) is the quantized value.

8. The method for intelligent highway lighting control integrating AI edge control according to claim 1, characterized in that: The energy efficiency optimization model takes into account the nonlinear characteristics of LED lamps. The optimization goal is to reduce the energy consumption of the lighting system while ensuring the lighting effect. The formula of the energy efficiency optimization model is as follows: Among them: a i 、b i and c i These three parameters represent the characteristics of the i-th luminaire; is the nonlinear part, which represents the relationship between energy consumption and the control input of the lamp setting, u i is the control input of the i-th luminaire; β is the smoothing coefficient, which is used to adjust the smoothness of the control input change. A larger β value helps reduce the volatility of the control input and prevent the impact of brightness fluctuations on the visual effect; Var(u) is the variance of the control input, which represents the fluctuation amplitude of the control input. In order to smooth the lighting changes, it reduces the rapid changes of the control input, improves the user experience and energy saving effect. The goal of the energy efficiency optimization model is to optimize the energy efficiency of the lighting system by adjusting the control input u i To minimize total energy consumption while ensuring smooth lighting and avoiding energy waste or undesirable lighting effects due to unstable control input.

9. The method for intelligent highway lighting control integrating AI edge control according to claim 1, characterized in that: The illumination gradient constraint is: Where: L i and L j Indicates the lighting brightness at different locations; It is the lighting brightness gradient, which indicates the brightness difference between two adjacent lamps. Controlling the lighting gradient helps to avoid sudden changes in lighting. E represents the connection between lamps, usually referring to adjacent lamps in the lamp layout. The constraints are: |L i -L j |≤0.3max(L i ,L j ) The brightness difference between adjacent lamps is required to not exceed 30% of the maximum brightness. This constraint ensures the smoothness of the lighting transition and avoids discomfort or interference to the driver caused by large brightness changes.

10. The method for intelligent highway lighting control integrating AI edge control according to claim 9, characterized in that: The triggering conditions of the emergency lighting are: L emergency =max(L normal ,L min +kv 2 ) Where: L emergency is the brightness of emergency lighting; L normal is the lighting brightness under normal conditions; L min is the minimum lighting brightness, ensuring that it will not fall below a certain minimum brightness in an emergency situation; κ is a coefficient related to vehicle speed, which represents the impact of speed on the brightness of emergency lighting; the faster the vehicle speed, the brighter the emergency lighting, to ensure that the lighting is high enough to ensure safety in high-speed situations; b is the vehicle speed.

Images