A multi-objective hierarchical depth optimization method for lighting control

Abstract

The application discloses a multi-target hierarchical depth optimization method for lighting control, and belongs to the technical field of lighting control. The specific steps of the optimization method are as follows: I, collecting and preprocessing various types of information at the edge nodes of the lighting area, and extracting semantic information describing real-time scenes from the preprocessed various types of information; the application significantly reduces the influence of sensing noise, missing data and heterogeneous sampling on decision-making, improves the accuracy and stability of environmental perception, enables the lighting control strategy to have both immediate response capability and forward-looking regulation capability, avoids the problem that the traditional fixed weight method is insufficient in adaptability under different times and different scenes, enables the energy saving, safety and comfort targets to dynamically balance according to the actual environment, effectively decomposes complex decision-making problems, improves the convergence speed and robustness of the strategy, simultaneously avoids the performance drop caused by scene extrapolation, and significantly improves the generalization capability of the system in a complex urban environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lighting control, and in particular to a multi-target hierarchical depth optimization method for lighting control. BACKGROUND

[0002] With the continuous advancement of smart cities and new infrastructure construction, urban lighting systems have gradually developed from traditional timing switch control to intelligent and adaptive direction. Modern urban lighting not only bears the basic function of night lighting, but also directly relates to traffic safety, public safety, urban image and energy consumption level. However, in the actual operation process, urban lighting control faces various complex and mutually restrictive target requirements, such as reducing energy consumption while ensuring road traffic and pedestrian safety, and prolonging equipment service life while improving lighting comfort. Most existing lighting control systems use rule-driven or single optimization target control methods, which are usually based on preset threshold, fixed time period strategy or simple perception feedback for adjustment. This kind of method has certain effect when the scene is single or the environment changes slowly, but when facing complex dynamic scenes such as traffic flow fluctuation, weather mutation, special event superposition, it is often difficult to balance multiple targets, and problems such as high energy consumption, insufficient lighting or frequent dimming often occur; Therefore, it is particularly important to invent a multi-target hierarchical depth optimization method for lighting control in a preset range.

[0003] In the existing multi-target hierarchical depth optimization method for lighting control, the influence of sensor noise, missing data and heterogeneous sampling on decision-making is high, which reduces the accuracy and stability of environmental perception. At the same time, the traditional lighting control strategy cannot have instant response ability while having forward-looking adjustment ability, and has the problem of insufficient adaptability in different times and different scenes. It cannot decompose complex decision-making problems, reduces the convergence speed and robustness of the strategy, and is easy to cause performance drop due to scene extrapolation; Therefore, we propose a multi-target hierarchical depth optimization method for lighting control. SUMMARY

[0004] The purpose of the present application is to solve the defects in the prior art and propose a multi-target hierarchical depth optimization method for lighting control.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0006] A multi-target hierarchical depth optimization method for lighting control, the specific steps of which are as follows:

[0007] I. Collect and preprocess various types of information at the edge nodes of the lighting area, and extract semantic information describing real-time scenes from the preprocessed various types of information;

[0008] II. Predict the traffic pattern, lighting demand and weather trend of future hours or days, and generate multi-mode candidate strategies;

[0009] III. Use the meta-strategy controller of the edge node to fuse real-time semantic information and multi-mode candidate strategies, and generate temporary sub-strategies when detecting unappeared scene combinations;

[0010] IV. Based on the control target weight output by the meta-strategy controller, activate the corresponding sub-strategy executor, and perform online adjustment on the newly generated temporary sub-strategy during execution;

[0011] V. Each lighting area shares meta-strategy parameters, and the strategy migration between different lighting areas is performed in an asynchronous update manner;

[0012] VI. Perform lighting control behavior, and synchronously evaluate the computing load, energy consumption and communication overhead of the edge computing node, and dynamically adjust each edge computing resource according to the evaluation result;

[0013] VII. Asynchronously update different levels of control parameters according to hours, minutes and seconds, and evaluate the performance of temporary sub-strategies after the scene ends.

[0014] As a further scheme of the present application, the specific steps of extracting the semantic information describing the real-time scene from the preprocessed various types of information in step I are as follows:

[0015] S1.1: Real-time acquisition of each observation data from various sensors at the edge node, adding corresponding time stamps to each observation data according to the sampling time of each sensor, and recording the sampling metadata, then sorting each group of observation data according to the sampling time from new to old, and aligning each type of observation data to a unified time grid through linear interpolation;

[0016] S1.2: Discrete filtering, outlier detection and single-point correction are performed on the unified observation data, and the systematic deviation between sensors is calibrated, then each descriptive feature is extracted from the preprocessed observation data, and a feature vector of illumination, traffic flow, people flow and weather is constructed for each time step, then the time information is periodically encoded, and the event information is converted into a sparse binary vector;

[0017] S1.3: A light filter is used to weight and fuse the feature vectors from illumination, traffic flow, people flow and weather at each time step, and the corresponding fusion state vector and confidence are generated, then the fusion state vector is mapped to a set of high-level semantic labels and the corresponding confidence by the edge node.

[0018] As a further scheme of the present application, the sampling metadata in S1.1 specifically includes sensor ID, sampling frequency, signal quality indicator, etc.

[0019] As a further scheme of the present application, the specific steps of predicting the future hour-level or day-level traffic pattern, lighting demand and weather trend in step II are as follows:

[0020] S2.1: Extract the observation sequence from the historical database according to the prediction target, construct a sliding historical window, establish a time series of traffic count sequence, lighting energy consumption sequence and weather element sequence for different data sources respectively, divide the time axis according to the target granularity, aggregate the data in the box according to the statistical rules, and record the box start time identifier to generate the historical feature matrix;

[0021] S2.2: Generate periodic encoding and exogenous variables from the start time of the time box, and generate corresponding lag difference features and moving average features based on the weather class historical data, then splice the generated periodic encoding and exogenous variables with the fusion state vector to form an extended feature vector for each historical box;

[0022] S2.3: Collect the extended feature vectors of each historical box and integrate them into a training set, construct an autoregressive baseline model, and input the training set into the autoregressive baseline model in batches, and obtain the autoregressive coefficient by least squares fitting. The autoregressive baseline model outputs the prediction results within a preset short period of time based on the generated autoregressive coefficient;

[0023] S2.4: Establish a corresponding long-term prediction model based on a small LSTM architecture, then input the training set into the long-term prediction model, the long-term prediction model processes each training data layer by layer through forward propagation, and outputs the prediction results of each training data through the gating mechanism, calculates the error value between the prediction results and the actual observation results, inputs the error value into the long-term prediction model based on the back propagation algorithm, calculates the gradient value of each layer parameter for the error value layer by layer, and adjusts each layer parameter through the Adam optimizer. Repeat the iteration training until the error value converges to a preset range, then input the latest extended feature vector into the long-term prediction model and perform forward propagation to input the prediction results within a preset long-term time;

[0024] S2.5: Model the weather data using a probability model with seasonal terms, and output the probability distribution of each weather index through Logistic regression, then map the weather index probability distribution results to binary or graded events according to the preset threshold, and combine the short-term prediction, long-term prediction and weather probability through weighted fusion method, and generate the final prediction sequence, and project the fused prediction sequence into the candidate strategy set through the strategy mapping rule or utility function.

[0025] As a further scheme of the present application, if the binning in S2.1 is by hour, each bin represents a 1-hour interval, and if it is by day, each bin represents 24 hours.

[0026] The periodic encoding and exogenous variables in S2.2 specifically include: hour angle encoding, day-of-week encoding, holiday flag, and known large event day.

[0027] As a further scheme of the present application, the specific steps of step III for fusing real-time semantic information and multi-mode candidate strategies by using the edge node meta-strategy controller and generating a temporary sub-strategy when an unemerged scene combination is detected are as follows:

[0028] S3.1: The edge node respectively vectorizes the high-level semantic label set and the candidate strategy set, and projects them to the same dimensional state space through linear mapping transformation to establish a fused state representation, and inputs the state representation to the meta-strategy controller;

[0029] S3.2: The meta-strategy controller calculates the relative importance scores of the current control targets through a lightweight evaluation network based on the received state representation, and then the meta-strategy controller converts the generated relative importance scores of each group into corresponding target weights, and through normalization operation, the sum of all target weights is 1;

[0030] S3.3: The meta-strategy controller generates the distribution characteristics of the emerged state combinations in the state space in real time through the built-in historical scene distribution model, and compares the current state representation with the historical distribution to determine whether it belongs to a known scene, and if the difference between the current state representation and the historical state set exceeds the preset threshold, it is considered that the state is an unemerged scene combination, and the temporary sub-strategy generation mechanism is triggered;

[0031] S3.4: When a new scene is detected, the meta-strategy controller calls the strategy generation network, and the strategy generation network takes the current state representation and target weight distribution as conditional input and outputs a set of strategy parameters to construct the corresponding temporary sub-strategy, and then the temporary sub-strategy is activated during the duration of the current scene.

[0032] As a further scheme of the present application, the specific steps of step IV for activating the corresponding sub-strategy executor and performing online adjustment on the newly generated temporary sub-strategy during the execution process are as follows:

[0033] S4.1: After the meta-strategy controller outputs the target weights of each target at the current time, the sub-strategy management module calculates the matching degree between each existing sub-strategy and the current target weight, and if the target preference of any sub-strategy is similar to the current weight distribution with a similarity higher than a preset threshold, the sub-strategy is directly activated, otherwise, the generated temporary sub-strategy is marked;

[0034] S4.2: The activated sub-strategy executor receives the current environment state information, while according to the current temporary sub-strategy target preference, the current state representation is cropped or reweighted to generate the internal state used at different times, and the action space is added with constraints according to the road level, lamp specifications and safety specifications, to limit the maximum brightness, minimum brightness and single change amplitude;

[0035] S4.3: The sub-strategy executor generates the corresponding control action based on the policy gradient method, and generates the corresponding action distribution or deterministic action value according to the internal state of the current temporary sub-strategy, and generates the transition smooth execution instruction by combining the current action and the actual execution action at the last time, and at the same time, the lamps belonging to the same logical lamp group are subjected to coordination constraints;

[0036] S4.4: After the action is executed, the sub-strategy executor collects the execution results in real time, weights and combines each execution result according to the target preference of the temporary sub-strategy, calculates the immediate reward signal, and adjusts each parameter in the temporary sub-strategy according to the immediate reward signal.

[0037] As a further scheme of the present application, the specific steps of step V for each lighting area to share the meta-strategy parameters are as follows:

[0038] S5.1: The lighting system is divided into multiple groups of logical regions, each logical region corresponds to an edge node, and an independent agent is deployed on each edge node, and each agent only uses the data of the corresponding logical region for perception, training and decision-making;

[0039] S5.2: Each regional agent updates the strategy parameters based on its own environmental characteristics after running locally for a period of time, and calculates the degree of change of the current agent parameters relative to the initial parameters or historical aggregated parameters, and if the degree of change is higher than a preset threshold, it indicates that new scene characteristics have appeared in the region;

[0040] S5.3: When any agent parameter changes more than a preset threshold, perform parameter aggregation operation, each regional agent uploads the local meta-strategy parameters to the federal coordination node, and then the federal coordination node aggregates the strategy parameters according to the contribution degree and sample size of each region to form a new global meta-strategy model;

[0041] S5.4: After completing the federal aggregation, each regional agent performs short-term trial operation in the local environment according to the preset trial time, and evaluates the performance change of the new strategy parameters in the region, and if the preset performance conditions are met, the new strategy parameters are confirmed to be effective, and are used as the formal running strategy of the current region.

[0042] As a further scheme of the present application, the specific steps of step VI for dynamically adjusting each edge computing resource according to the evaluation results are as follows:

[0043] S6.1: While the lighting control task is being executed, the edge computing node continuously monitors its own running state, wherein the running state includes processor usage, memory occupation, unit time energy consumption, and sending and receiving rates of the network interface, each monitoring data is collected according to a unified sampling period, and is mapped into a resource index;

[0044] S6.2: The resource index is normalized by a Max-Min normalization method, and then each edge node constructs a joint resource consumption evaluation function based on the corresponding normalized resource index, and calculates the comprehensive resource pressure index of each edge node at different times through the function, and judges the current node load state according to the comprehensive resource pressure index;

[0045] S6.3: When the comprehensive resource pressure index exceeds a preset threshold, the execution frequency of the computation-intensive task is preferentially reduced, and each edge node reorders the priority of each type of task executed by itself, and then adjusts the running parameters and executes according to the task execution frequency and the priority reordering result, and continuously monitors the node resource state after the adjustment.

[0046] Compared with the prior art, the beneficial effects of the present application are:

[0047] The multi-objective hierarchical deep optimization method for lighting control collects and pre-processes multiple types of sensor data in real time through the edge node, then extracts illumination, traffic flow, pedestrian flow and weather features, combines time period coding and event information for feature fusion, generates a fusion state vector and a confidence, and maps it into a high-level semantic label, constructs a multi-source time series and a sliding window feature based on historical data, uses an autoregressive model for short-term prediction, an LSTM for long-term prediction, and combines a probability model to output a weather event probability, fuses to form a final prediction sequence and maps it into a candidate strategy, then the fusion state and the candidate strategy are jointly constructed into a unified state representation input meta-strategy controller, which adaptively generates target weights, identifies new scenes and generates temporary sub-strategies as needed, the sub-strategies output smooth control actions within the constraint action space and adjust parameters online, and independent agents are deployed in multiple regions, aggregate meta-strategy parameters through federated learning, significantly reduce the influence of sensing noise, missing data and heterogeneous sampling on decision-making, improve the accuracy and stability of environmental perception, make the lighting control strategy have both immediate response capability and forward-looking adjustment capability, avoid the problem of insufficient adaptability of traditional fixed weight methods in different times and different scenes, dynamically balance energy saving, safety and comfort according to the actual environment, effectively decompose complex decision-making problems, improve the convergence speed and robustness of the strategy, and avoid performance degradation caused by scene extrapolation, significantly improve the generalization ability of the system in complex urban environments. BRIEF DESCRIPTION OF DRAWINGS

[0048] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are meant to explain the present application and are not intended to limit the application.

[0049] Figure 1 A flow chart of a multi-objective hierarchical depth optimization method for lighting control is proposed. DETAILED DESCRIPTION

[0050] Embodiment 1, with reference to Figure 1 A multi-objective hierarchical depth optimization method for lighting control, the specific steps of which are as follows:

[0051] Collect and preprocess various types of information at the edge nodes of the lighting area, and extract semantic information describing the real-time scene from the preprocessed various types of information.

[0052] Specifically, at the edge node, real-time observation data is collected from various sensors, time stamps are added to each observation data according to the sampling time of each sensor, and sampling metadata is recorded. Then, each group of observation data is sorted from new to old according to the sampling time, and various types of observation data are aligned to a unified time grid through linear interpolation. Discrete filtering, outlier detection and single-point correction are performed on the unified various types of observation data, and systematic bias between sensors is calibrated. Then, each descriptive feature is extracted from the preprocessed various types of observation data, a feature vector of illumination, traffic flow, people flow and weather is constructed for each time step, and then the time information is periodically encoded, and the event information is converted to a sparse binary vector. A light filter is used to weight and fuse the feature vectors from illumination, traffic flow, people flow and weather at each time step, and to correct the confidence, to generate the corresponding fusion state vector and confidence. Then, the edge node maps the fusion state vector to a set of high-level semantic labels and the corresponding confidence.

[0053] It should be noted that the sampling metadata specifically includes sensor ID, sampling frequency, signal quality index, etc.

[0054] Predict the traffic pattern, lighting demand and weather trend in the future hour or day level, and generate multi-mode candidate strategies.

[0055] Specifically, the observation sequence is extracted from the historical database according to the prediction target, a sliding historical window is constructed, the traffic count sequence, the lighting energy consumption sequence and the weather element sequence are respectively established for different data sources, the time axis is binned according to the target granularity, the data in the bin is aggregated according to the statistical rule, and the bin start time identifier is recorded to generate a historical feature matrix, the periodic encoding and the exogenous variable are generated from the start time of the time bin, the corresponding lag difference features and moving average features are generated based on the weather class historical data, then the generated periodic encoding and exogenous variable are spliced with the fusion state vector to form an extended feature vector of each historical bin, the extended feature vectors of the historical bins are collected and integrated into a training set, an autoregressive baseline model is constructed, the training set is input into the autoregressive baseline model in batches, and the autoregressive coefficients are obtained by least squares fitting, the autoregressive baseline model outputs the prediction result in a preset short-term time based on the generated autoregressive coefficients, a corresponding long-term prediction model is established based on a small LSTM architecture, then the training set is input into the long-term prediction model, the long-term prediction model processes each training data layer by layer through forward propagation, and outputs the prediction result of each training data through the gating mechanism, the error value between the prediction result and the actual observation result is calculated, the error value is input into the long-term prediction model based on the back propagation algorithm, the gradient value of the error value to each layer parameter is calculated layer by layer, and each layer parameter is adjusted by the Adam optimizer, the training is repeated and iterated until the error value converges to a preset range, then the latest extended feature vector is input into the long-term prediction model, and the prediction result in a preset long-term time is output through forward propagation, a probability model with seasonal terms is used to model the weather data, and the probability distribution of each weather index is output through Logistic regression, then each weather index probability distribution result is mapped to a binary or graded event according to a preset threshold, and the short-term prediction, the long-term prediction and the weather probability are combined by a weighted fusion method, and the final prediction sequence is generated, and the fused prediction sequence is projected into a candidate strategy set through a strategy mapping rule or an utility function.

[0056] In this embodiment, if the binning is per hour, each bin represents a 1-hour interval, and if the binning is per day, each bin represents 24 hours; the periodic encoding and the exogenous variable specifically include: hour angle encoding, day of week encoding, holiday flag, and known large event day flag.

[0057] The meta-strategy controller of the edge node fuses real-time semantic information and multi-mode candidate strategies, and generates a temporary sub-strategy when an unappeared scene combination is detected.

[0058] Specifically, the edge node respectively vectorizes the high-level semantic label set and the candidate strategy set, and projects them to the same dimensional state space through linear mapping transformation to establish a fused state representation, and inputs it to the meta-strategy controller. The meta-strategy controller calculates the relative importance scores of the current control targets through a lightweight evaluation network based on the received state representation. Then, the meta-strategy controller converts the generated relative importance scores of each group into corresponding target weights, and normalizes the sum of all target weights to 1. The meta-strategy controller generates the distribution characteristics of the existing state combinations in the state space in real time through the built-in historical scene distribution model, and compares the current state representation with the historical distribution to determine whether it belongs to a known scene. If the difference between the current state representation and the historical state set exceeds the preset threshold, it is considered that the state is a new scene combination, and the temporary sub-strategy generation mechanism is triggered. When a new scene is detected, the meta-strategy controller calls the strategy generation network, which takes the current state representation and target weight distribution as input conditions and outputs a set of strategy parameters to construct the corresponding temporary sub-strategy. Then, the temporary sub-strategy is activated during the duration of the current scene.

[0059] Embodiment 2, refer to Figure 1 A multi-objective hierarchical deep optimization method for lighting control, the specific steps of the optimization method are as follows:

[0060] Based on the control target weights output by the meta-strategy controller, the corresponding sub-strategy executor is activated, and the newly generated temporary sub-strategy is adjusted online during execution.

[0061] Specifically, after the meta-strategy controller outputs the target weights of the current time, the sub-strategy management module calculates the matching degree between the existing sub-strategies and the current target weights. If the target preference of any sub-strategy is similar to the current weight distribution with a similarity higher than a preset threshold, the sub-strategy is directly activated. Otherwise, the generated temporary sub-strategy is marked. The activated sub-strategy executor receives the current environment state information, and simultaneously performs cropping or re-weighting on the current state representation according to the target preference of the current temporary sub-strategy to generate internal states used at different times. Constraints are added to the action space according to the road level, lamp specifications and safety specifications to limit the maximum brightness, minimum brightness and single change amplitude. The sub-strategy executor generates the corresponding control action based on the policy gradient method, and generates the corresponding action distribution or deterministic action value according to the internal state of the current temporary sub-strategy. The transition-smoothed execution instruction is generated by combining the current action with the actual execution action at the last time. At the same time, coordination constraints are applied to the lamps belonging to the same logical lamp group. After the action is executed, the sub-strategy executor collects the execution results in real time, weights and combines the execution results according to the target preference of the temporary sub-strategy, calculates the immediate reward signal, and adjusts the parameters in the temporary sub-strategy according to the immediate reward signal.

[0062] Each lighting area shares the meta-strategy parameter, and the inter-area asynchronous update method is used for policy migration between different lighting areas.

[0063] Specifically, the lighting system is divided into multiple groups of logical areas, each logical area corresponds to an edge node, and an independent agent is deployed on each edge node, and each agent only uses the data of the corresponding logical area for perception, training and decision-making. After each regional agent runs locally for a period of time, it will update the policy parameters based on its environmental characteristics, and calculate the degree of change of the current agent parameters relative to the initial parameters or historical aggregated parameters. If the degree of change is higher than the preset threshold, it means that there is a new scene feature in the region. When the parameter of any agent changes more than the preset threshold, perform parameter aggregation operation, each regional agent uploads the local meta-strategy parameter to the federal coordination node, and then the federal coordination node aggregates the policy parameters according to the contribution and sample size of each region to form a new global meta-strategy model. After completing the federal aggregation, each regional agent will update the policy in the local environment according to the preset trial time for short-term trial operation, and evaluate the performance change of the new policy parameters in the region. If it meets the preset performance conditions, it is confirmed that the new policy parameters are effective, and they are used as the formal running strategy of the current region.

[0064] The lighting control behavior is executed, and the computing load, energy consumption and communication overhead of the edge computing node are synchronously evaluated, and the computing resources of each edge are dynamically adjusted according to the evaluation results.

[0065] Specifically, while the lighting control task is being executed, the edge computing node continuously monitors its own running state, including processor usage, memory occupancy, unit time energy consumption, and network interface sending and receiving rate. The monitoring data is collected according to a unified sampling period and mapped to resource indicators. The resource indicators are normalized by the Max-Min normalization method. Based on the corresponding normalized resource indicators, each edge node constructs a joint resource consumption evaluation function, and calculates the comprehensive resource pressure index of each edge node at different times through the function. According to the comprehensive resource pressure index, the current node load state is determined. When the comprehensive resource pressure index exceeds the preset threshold, the execution frequency of the computation-intensive task is reduced first. At the same time, each edge node reorders the priority of the tasks it executes. Then, according to the task execution frequency and the priority reordering result, each edge node adjusts the running parameters and executes in real time. After adjustment, the node resource state is continuously monitored.

[0066] The different levels of control parameters are asynchronously updated at the hour level, minute level and second level, and the performance of the temporary sub-strategy is evaluated and stored after the scene ends.

Claims

1. A multi-objective hierarchical depth optimization method for lighting control, characterized in that, The specific steps of this optimization method are as follows: I. Collect and preprocess various types of information at the edge nodes of the lighting area, and extract semantic information describing the real-time scene from the preprocessed information. II. Predict future hourly or dayly traffic patterns, lighting demands, and weather trends to generate multi-mode candidate strategies; III. Utilize the meta-policy controller of edge nodes to fuse real-time semantic information with multi-pattern candidate policies, and generate temporary sub-policies when a scenario combination that does not appear is detected; the specific steps are as follows: S3.1: Edge nodes vectorize and encode the high-level semantic label set and the candidate policy set respectively, and project them onto the state space of the same dimension through linear mapping transformation to establish the fused state representation, and input it into the meta-policy controller; S3.2: Based on the received state representation, the meta-policy controller calculates the relative importance score of each control objective through a lightweight evaluation network. Then, the meta-policy controller converts the generated relative importance scores into corresponding objective weights and performs a normalization operation to sum all objective weights to 1. S3.3: The meta-policy controller generates the distribution characteristics of the already occurring state combinations in the state space in real time through the built-in historical scene distribution model, and compares the current state representation with the historical distribution to determine whether it belongs to a known scene. If the difference between the current state representation and the historical state set exceeds a preset threshold, the state is considered to be a scene combination that has not appeared, and a temporary sub-policy generation mechanism is triggered. S3.4: When a new scene is detected, the meta-policy controller calls the policy generation network. The policy generation network takes the current state representation and the target weight distribution as input conditions and outputs a set of policy parameters to construct the corresponding temporary sub-policy. The temporary sub-policy is then activated during the duration of the current scene and is adjusted in real time using online data. IV. Based on the control target weights output by the meta-policy controller, activate the corresponding sub-policy executor and adjust the newly generated temporary sub-policy online during the execution process; V. Each lighting zone shares the meta-strategy parameters, and at the same time, an asynchronous update method is used between different lighting zones to migrate strategies. VI. Execute lighting control behavior and simultaneously evaluate the computing load, energy consumption and communication overhead of edge computing nodes, and dynamically adjust each edge computing resource based on the evaluation results; VII. Asynchronously update control parameters at different levels by hour, minute, and second, respectively, and evaluate and store the performance of temporary sub-strategies after the scenario ends.

2. The multi-objective layered depth optimization method for lighting control according to claim 1, characterized in that, The specific steps for extracting semantic information describing the real-time scene from the preprocessed information in step I are as follows: S1.1: Collect observation data from various sensors in real time at edge nodes, add corresponding timestamps to each observation data according to the sampling time of each sensor, and record sampling metadata. Then, sort each group of observation data from newest to oldest according to the sampling time, and then align various observation data to a unified time grid through linear interpolation. S1.2: Discrete filtering, outlier detection and single-point correction are performed on the unified observation data of various types, and the systematic deviation between sensors is calibrated. Then, descriptive features are extracted from the preprocessed observation data of various types, and feature vectors of illumination, traffic flow, pedestrian flow and weather are constructed for each time step. Then, the time information is periodically encoded, and the event information is converted into a sparse binary vector. S1.3: A lightweight filter is used to perform weighted fusion and confidence correction on the feature vectors from illumination, traffic flow, pedestrian flow and weather at each time step to generate the corresponding fused state vector and confidence. Then, the fused state vector is mapped to a set of high-level semantic labels and corresponding confidence through edge nodes.

3. The multi-objective layered depth optimization method for lighting control according to claim 2, characterized in that, The specific steps for predicting future hourly or daily traffic patterns, lighting needs, and weather trends, as described in Step II, are as follows: S2.1: Extract observation sequences from the historical database according to the prediction target, construct a sliding historical window, establish traffic count sequence, lighting energy consumption sequence and weather element sequence for different data sources, divide the time axis into bins according to the target granularity, aggregate the data in the bins according to statistical rules, and record the bin start time identifier to generate a historical feature matrix. S2.2: Generate periodic codes and exogenous variables from the start time of the time bin, and generate corresponding lag difference features and moving average features based on historical weather data. Then, concatenate the generated periodic codes and exogenous variables with the fused state vector to form the extended feature vector of each historical bin. S2.3: Collect the extended feature vectors of each historical bin and integrate them into a training set to build an autoregressive baseline model. Input the training set into the autoregressive baseline model in batches and obtain the autoregressive coefficients through least squares fitting. Based on the generated autoregressive coefficients, the autoregressive baseline model outputs the prediction results within a preset short period of time. S2.4: A corresponding long-term prediction model is established based on a small LSTM architecture. The training set is then input into the long-term prediction model. The long-term prediction model processes each training data layer by layer through forward propagation and outputs the prediction results of each training data through a gating mechanism. The error value between the prediction result and the actual observation result is calculated. The error value is input into the long-term prediction model based on the backpropagation algorithm. The gradient value of the error value with respect to the parameters of each layer is calculated layer by layer. The parameters of each layer are adjusted through the Adam optimizer. The training is iterated repeatedly until the error value converges to the preset range. Then, the latest extended feature vector is input into the long-term prediction model, and the prediction results within the preset long-term time are input through forward propagation. S2.5: The weather data is modeled using a probability model with seasonal terms, and the probability distribution of each weather indicator is output through Logistic regression. Then, the probability distribution results of each weather indicator are mapped to binary or graded events according to a preset threshold. The short-term forecast, long-term forecast and weather probability are then merged through a weighted fusion method to generate the final forecast sequence. The fused forecast sequence is then projected into a set of candidate strategies through policy mapping rules or utility functions.

4. The multi-objective layered depth optimization method for lighting control according to claim 1, characterized in that, The specific steps for activating the corresponding sub-policy executor in step IV, and for making online adjustments to the newly generated temporary sub-policy during execution, are as follows: S4.1: After the meta-policy controller outputs the target weights at the current moment, the sub-policy management module calculates the matching degree between each existing sub-policy and the current target weight. If the similarity between the target preference of any sub-policy and the current weight distribution is higher than a preset threshold, the sub-policy is activated directly; otherwise, the generated temporary sub-policy is marked. S4.2: The activated sub-policy executor receives the current environmental state information and, based on the current temporary sub-policy target preference, trims or reweights the current state representation to generate internal states used at different times. It also adds constraints to the action space based on road grade, lighting specifications, and safety regulations, limiting the maximum brightness, minimum brightness, and single change range. S4.3: The sub-policy executor generates corresponding control actions based on the policy gradient method, and generates corresponding action distributions or deterministic action values ​​according to the internal state of the current temporary sub-policy. It also combines the current action with the actual execution action of the previous moment to generate a smooth-transition execution instruction. At the same time, it applies coordination constraints to lamps belonging to the same logical lamp group. S4.4: After the action is executed, the sub-policy executor collects the execution results in real time. According to the target preference of the temporary sub-policy, it performs weighted combination of each execution result, calculates the instant reward signal, and adjusts the various parameters in the temporary sub-policy according to the instant reward signal.

5. The multi-objective layered depth optimization method for lighting control according to claim 1, characterized in that, The specific steps for sharing the meta-strategy parameters among the lighting areas in step V are as follows: S5.1: Divide the lighting system into multiple logical regions, each corresponding to an edge node, and deploy an independent agent on each edge node. At the same time, each agent only uses the data of the corresponding logical region for perception, training and decision-making. S5.2: After running locally for a period of time, each regional agent will update the policy parameters based on its own environmental characteristics, and at the same time calculate the degree of change of the current agent parameters relative to the initial parameters or historical aggregated parameters. If the degree of change is higher than the preset threshold, it indicates that new scene features have appeared in the region. S5.3: When the parameter change of any agent exceeds the preset threshold, the parameter aggregation operation is performed. Each regional agent uploads its local meta-policy parameters to the federated coordination node. Then, the federated coordination node performs weighted aggregation of each policy parameter according to the contribution of each region and the sample size to form a new global meta-policy model. S5.4: After completing federated aggregation, each regional agent will conduct a short-term trial run of the updated policy in the local environment according to the preset trial period, and evaluate the performance changes of the new policy parameters in the region. If the preset performance conditions are met, the new policy parameters are confirmed to be effective and used as the official running policy for the current region.

6. The multi-objective hierarchical depth optimization method for lighting control according to claim 1, characterized in that, The specific steps for dynamically adjusting each edge computing resource based on the evaluation results, as described in step VI, are as follows: S6.1: While the lighting control task is being executed, the edge computing node continuously monitors its own operating status, including processor utilization, memory usage, energy consumption per unit time, and the sending and receiving rates of the network interface. All monitoring data are collected according to a uniform sampling period and mapped to resource indicators. S6.2: The resource indicators are normalized using the Max-Min normalization method. Then, each edge node constructs a joint resource consumption evaluation function based on the corresponding normalized resource indicators. The comprehensive resource pressure index of each edge node at different times is calculated using this function. The load status of the current node is determined based on the comprehensive resource pressure index. S6.3: When the comprehensive resource pressure index exceeds the preset threshold, the execution frequency of computationally intensive tasks is reduced first. At the same time, each edge node re-prioritizes the various tasks it executes. Then, based on the task execution frequency and the priority re-priority results, each edge node adjusts its running parameters and executes them in real time. After adjustment, the node resource status is continuously monitored.

Images