Multi-node self-adaptive dimming intelligent street lamp centralized control system

Abstract

The invention relates to the technical field of smart city lighting control, and provides a multi-node adaptive dimming smart street lamp centralized control system. The system constructs an overall illumination state model in space and time dimensions by accessing multi-node intelligent street lamp operation data and external state data related to illumination requirements, and the overall illumination state model is used for uniformly representing the illumination distribution state in a road or an area and dynamically updating the illumination distribution state. Based on the overall lighting state model, the system performs situation analysis on lighting demand changes and evaluates uncertainty of data, state models or analysis results to constrain dimming decision generation. According to the system, multi-node cooperative dimming is achieved in a centralized control mode, self-adaptive optimization is conducted on a state model and a dimming strategy in combination with operation feedback, comprehensive balance between illumination quality and energy consumption is achieved, and therefore the stability, robustness and energy utilization efficiency of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of smart city lighting control technology, specifically to a centralized control system for smart streetlights with multi-node adaptive dimming. Background Technology

[0002] With the advancement of smart city and new urban infrastructure construction, smart streetlights, as an important component of urban public lighting systems, are gradually evolving from traditional timed switching and fixed brightness control methods to intelligent control based on environmental perception and communication networks. Existing smart streetlight systems typically deploy light sensors, human or vehicle sensing devices, and wireless communication modules to automatically adjust the lighting status of individual streetlights or local road sections, thereby reducing energy consumption while meeting lighting needs.

[0003] In existing technologies, common smart street light control methods mainly include automatic dimming based on ambient light thresholds, rule-driven dimming based on traffic flow or pedestrian activity, and time-sharing control strategies based on preset time periods. These methods typically use a single street light node or a small number of adjacent nodes as control units, independently determining the dimming strategy based on local sensing data. The control logic is relatively simple and easy to implement.

[0004] However, with the expansion of smart street light deployment and the increasing complexity of application scenarios, the aforementioned control methods based on individual lights or local rules have gradually revealed several shortcomings. On the one hand, individual street light nodes can only perceive local environmental information and lack awareness of the overall lighting status of the road or area, easily leading to problems such as inconsistent brightness between different street lights, and areas that are too bright or too dim, affecting driving safety and pedestrian visual comfort. On the other hand, independent decision-making by each street light node makes it difficult to achieve global optimization between lighting quality and energy consumption at the system level.

[0005] To address these issues, some existing technologies have proposed centralized or semi-centralized smart street light control systems. These systems aggregate operational data from multiple street light nodes through a central control platform and uniformly issue dimming commands. While these systems improve coordination among multiple street lights to some extent, their control strategies are typically based on discrete rules, simple statistical indicators, or empirical thresholds. They lack a systematic model of the overall lighting distribution and struggle to accurately reflect the continuous changes in lighting conditions across space and time.

[0006] Furthermore, existing smart street light control systems generally fail to explicitly model and evaluate the uncertainties of perceived data and analysis results during the decision-making process. Due to factors such as changes in ambient light, traffic flow fluctuations, sensor noise, communication delays, or data gaps, the external state and operational data acquired by the system are often uncertain. In the absence of an uncertainty assessment mechanism, dimming decisions are easily affected by occasional disturbances or measurement errors, leading to frequent dimming, decreased lighting stability, and even increased equipment wear and maintenance costs.

[0007] Meanwhile, most existing smart street light control solutions focus on single-node or simple linkage control, lacking control mechanisms for multi-node collaborative optimization. When multiple street light nodes participate in dimming simultaneously, it is difficult to coordinate the control objectives and constraints between different nodes, making it difficult to achieve coordinated dimming and energy consumption optimization at the regional scale.

[0008] In summary, existing smart street light control technologies still have shortcomings in areas such as unified modeling of overall lighting status, multi-node collaborative dimming, uncertainty assessment, and robust decision-making. There is an urgent need for a centralized smart street light control scheme that can uniformly represent lighting status at the system level, comprehensively analyze changes in lighting demand, and achieve multi-node collaborative dimming in uncertain environments. Summary of the Invention

[0009] To address the common problems in existing smart street light control systems, such as the lack of unified modeling of the overall lighting status, insufficient coordination due to independent dimming of multiple street light nodes, and low stability and reliability of dimming decisions under incomplete environmental perception data or noise conditions, it is necessary to propose a centralized control technology solution for smart street lights that can uniformly represent the lighting status at the system level, comprehensively analyze changes in lighting demand, and achieve multi-node collaborative dimming under uncertain environmental conditions.

[0010] To address this, a multi-node adaptive dimming intelligent street light centralized control system is proposed, comprising:

[0011] The data interface layer is used to receive and aggregate multi-source operation data from multi-node smart streetlights, as well as external status data related to lighting needs.

[0012] The data interface layer is used to receive and aggregate multi-source operation data and external status data from multi-node smart streetlights and external data sources, and to preprocess, verify and cache them.

[0013] The data interface layer can receive reported data from street light nodes through message queues or publish / subscribe mechanisms. The reported data may include: node identifier, timestamp, current brightness parameter or drive duty cycle parameter, energy consumption parameter, operating status information, and fault status information. External status data can be accessed through API interfaces, database interfaces, or streaming push methods, and has an update cycle or triggering mechanism corresponding to the data source.

[0014] To adapt to data sources with different sampling frequencies, the data interface layer can map data from different sources to a unified time base, forming a time-aligned data sequence; and can mark data missing due to missing fields, abnormal range values, communication delays, or packet loss. For missing or supplemented data, a "confidence flag" or "uncertainty flag" can be attached for subsequent use by the uncertainty assessment module.

[0015] The lighting state construction module is used to construct an overall lighting state model in the spatial and temporal dimensions of a road or region based on the multi-source operational data and external state data. The overall lighting state model serves as an intermediate state representation of the system, used to uniformly characterize the lighting distribution state at different spatial locations, and as a state input for multi-node collaborative dimming control.

[0016] The lighting situation analysis module is used to analyze and infer changes in lighting demand based on the overall lighting situation model.

[0017] The uncertainty assessment module is used to assess the uncertainty of the external state data, the overall lighting state model, or the lighting situation analysis results, and generate uncertainty assessment results.

[0018] The dimming decision and centralized control module is used to generate and implement a collaborative dimming control strategy for multiple smart street light nodes based on the comprehensive lighting situation analysis results and the uncertainty assessment results.

[0019] The feedback and optimization module is used to obtain operational feedback information after dimming control is executed, and to update or adjust the system's state model or dimming control strategy based on the operational feedback information.

[0020] The multi-source operation data includes at least the node identifier, spatial location information, current brightness parameter or drive duty cycle parameter, energy consumption parameter and operation status information of the smart street light node; the external status data includes at least one of ambient light data, traffic flow data, personnel activity data or weather visibility data.

[0021] The overall lighting state model is a light field model used to describe the distribution of illuminance or brightness values ​​at different locations within a road or area.

[0022] The light field model is continuously updated over time to reflect the changing trend of the lighting state in the time dimension and to serve as the state input for dimming decisions.

[0023] The lighting state construction module can estimate the illuminance / luminance of spatial units based on the brightness parameters, node position parameters, and luminaire light distribution information of street light nodes, and can further integrate the contribution of ambient light to obtain the overall lighting distribution. The luminaire light distribution information, spatial propagation parameters, etc., can be obtained through simulation calibration, on-site test calibration, or historical data fitting.

[0024] The light field model is used to characterize the illuminance or brightness distribution within a road or area. The light field model employs either a continuous spatial representation or a discrete grid representation, that is, dividing the road longitudinally into multiple road segment units and laterally into multiple lanes or functional areas, resulting in a set of spatial units used to describe the lighting distribution.

[0025] The light field model can be continuously updated over time to reflect the changing trend of the illumination state in the time dimension. The update method includes at least one of periodic updates and event-triggered updates: periodic updates refresh the global state at a predetermined time step; event-triggered updates can trigger local or global state reconstruction when a node failure, a step change in brightness, or a significant change in the external state is detected.

[0026] The lighting situation analysis module is used to analyze and infer changes in lighting demand based on the overall lighting state model. The module can construct a causal relationship model among factors influencing lighting demand to analyze the triggering conditions, scope, or degree of impact of changes in lighting demand.

[0027] The causal relationship model can employ a structural causal model, a causal relationship diagram, or other causal representation methods. The causal relationship edges can be configured with causal influence intensity parameters and time delay parameters to characterize the magnitude of the impact of external state changes on lighting demand and the response time lag. Influencing factors may include ambient light, traffic flow, population activity density, weather visibility, time period, and current illuminance level.

[0028] The lighting situation analysis module analyzes the triggering conditions, scope of influence, or degree of influence of changes in lighting demand by constructing a causal relationship model among the factors influencing lighting demand.

[0029] The causal relationship model is represented in the form of a structural causal model or a causal relationship graph, wherein the causal relationship edge configuration includes causal influence coefficients and time delay parameters.

[0030] The lighting situation analysis module can infer the target illuminance requirement at the current moment based on the current observation value, and can combine the short-term prediction results of the external state to predict the trend of lighting demand in the future, so as to support advance dimming.

[0031] The uncertainty assessment result is at least one of a confidence parameter, an uncertainty metric, or a probability distribution form.

[0032] The uncertainty assessment module quantifies uncertainty based on Bayesian inference methods, the statistical variance of prediction results, or the degree of discrepancy between prediction results from multiple models.

[0033] Sources of uncertainty may include, but are not limited to: sensor noise, data loss due to communication delays and packet loss, inaccuracy of external data, model parameter calibration errors, model structure biases, and inherent uncertainties arising from prediction. The uncertainty assessment result may be at least one of a confidence parameter, an uncertainty metric, or a probability distribution form.

[0034] The uncertainty assessment results can serve as an important constraint input for dimming decisions: when the uncertainty is high, the system can adopt a robust or conservative dimming strategy to preserve illuminance margin; when the uncertainty is low, the system can adopt a refined dimming strategy that is closer to the predicted demand to improve energy-saving effect.

[0035] When implementing collaborative dimming control, the dimming decision and centralized control module introduces at least one of lighting target constraints or energy consumption constraints. The dimming decision and centralized control module includes multiple functional software agents. Based on sharing the overall lighting state model, different software agents determine their respective dimming parameters through collaboration or game theory.

[0036] The feedback and optimization module adaptively updates the causal relationship model, uncertainty assessment results, or dimming control strategy based on operational feedback information.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] I. This invention uses an overall lighting state model as the unified state space of the system, uncertainty assessment as a decision credibility constraint mechanism, and multi-node collaborative dimming as a centralized control execution method. The overall lighting state model is no longer merely a local perception result or a basis for single-lamp control, but rather a unified state representation of a multi-node collaborative dimming system. Brightness, energy consumption, and operational status information from different smart street light nodes, as well as external state data such as ambient light, traffic flow, pedestrian activity, and weather visibility, are mapped onto a unified space-time lighting state model. This creates a consistent understanding of the overall lighting distribution of the road or area at the system level, avoiding the problems of uneven lighting, unreasonable energy consumption allocation, or dimming conflicts caused by independent decisions made by each street light node based on local sensing information in existing technologies.

[0039] Second, uncertainty assessment is no longer merely a tool for data quality analysis or prediction error description, but directly participates in the dimming decision generation process. While constructing the overall lighting state model and conducting lighting situation analysis, the system quantitatively assesses the uncertainty of external state data, state modeling errors, and the reliability of situation analysis results. The uncertainty assessment results serve as a crucial constraint on dimming decisions, enabling the system to proactively adopt robust dimming strategies even when sensing incomplete information or experiencing drastic environmental changes, avoiding frequent dimming or excessive energy-saving behavior due to misjudgments. The multi-node collaborative dimming mechanism, based on a shared and unified lighting state model and constrained by uncertainty assessment, achieves centralized and coordinated dimming control. This realizes a shift from "single-point passive adjustment" to "regional collaborative and adaptive adjustment."

[0040] Third, the overall lighting state model represents the illuminance or brightness state at different spatial locations as a continuous spatial-temporal distribution. This allows the system to assess lighting from the perspective of the entire area or road, rather than relying solely on the local brightness parameters of individual lamps, thus significantly improving the overall consistency and visual comfort of lighting control. This unified state model serves as a shared state input for multiple dimming decision units, providing a consistent state cognitive basis for dimming decisions across different road sections and functional areas. This enables adjacent street light nodes to perform collaborative control under a unified state semantic, avoiding brightness gaps or localized over-brightness or under-brightness phenomena that occur in traditional distributed control.

[0041] Meanwhile, the overall lighting state model has spatial continuity and temporal evolution characteristics, providing a basis for the analysis and prediction of lighting demand change trends. Compared with dimming methods based on discrete thresholds, it can more finely depict the gradual process of lighting demand and improve the smoothness and stability of dimming strategies.

[0042] Fourth, this invention introduces uncertainty assessment into the dimming decision-making process, enabling the system to identify the credibility of changes in lighting demand. By quantifying the uncertainty of sensor data noise, external state fluctuations, and model prediction errors, the system can distinguish between stable changes and occasional disturbances, avoiding over-response to short-term anomalies or measurement errors, thereby reducing unnecessary dimming operations and equipment wear.

[0043] The system can adopt conservative or robust dimming strategies under high uncertainty conditions. When environmental perception data is missing, communication is delayed, or prediction results are significantly divergent, the system constrains the dimming amplitude or adjusts the dimming rhythm based on uncertainty assessment results, improving the safety and reliability of the system in complex environments. By incorporating uncertainty assessment results into the decision-making process in the form of confidence parameters, probability distributions, or uncertainty metrics, the system achieves quantitative management of dimming risks, improving the interpretability and controllability of dimming decisions.

[0044] Through the aforementioned closed-loop mechanism, this invention organically combines data acquisition, lighting state modeling, situational analysis, uncertainty assessment, collaborative dimming decision-making, and operational feedback and optimization to construct a complete closed-loop control system. The operational feedback information acquired after dimming is executed is used not only to adjust current dimming parameters but also to continuously update the overall lighting state model, uncertainty assessment results, and dimming control strategy, enabling the system to adapt to different road scenarios, traffic conditions, and seasonal changes. Furthermore, this invention achieves adaptive optimization of the lighting control strategy based on the operating environment and historical feedback, improving the system's robustness, energy efficiency, and lighting service quality during long-term operation. Attached Figure Description

[0045] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0046] Figure 1 This is a diagram of the centralized control system architecture for a smart street light with multi-node adaptive dimming. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0049] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0050] Example 1

[0051] See Figure 1 This embodiment provides a specific implementation of a multi-node adaptive dimming smart street light centralized control system. The system is deployed in an urban road lighting scenario, managing a main road approximately 2 kilometers long, with a total of 120 smart street light nodes. Each street light node is equipped with an LED light source module, a local control unit, multiple sensors (including ambient light sensors and infrared motion sensors), a communication module, and an energy metering unit. The street light nodes maintain connectivity with the central control system via a wireless Mesh network or LoRa communication protocol. The central control system is deployed in the cloud or on an edge server, responsible for aggregating operational data across the entire road segment, constructing the overall lighting status, analyzing lighting demand trends, assessing decision uncertainties, and generating collaborative dimming strategies.

[0052] Each street light node has a unique node ID, and its spatial coordinates (latitude and longitude or relative position), installation height, and pole spacing are recorded during system initialization. During operation, the street light nodes continuously report dynamic operating data, including the current drive duty cycle (corresponding to brightness output), actual power consumption, light source operating temperature, and fault status indicators. In addition, the system connects to external data sources, including real-time weather visibility data from a meteorological service platform, traffic flow data from a traffic management system, and pedestrian activity density data from a video surveillance system.

[0053] As the underlying foundational module of the system, the data interface layer is responsible for collecting raw data from distributed street light nodes and external data sources and performing preprocessing.

[0054] Multi-source data access. The system receives real-time data packets from each street light node via message queues (such as MQTT or Kafka). Each data packet contains a node identifier, timestamp, brightness parameters (current duty cycle, range 0-100%), energy consumption parameters (instantaneous power, unit W), and operating status flags (normal / fault / communication anomaly). For external data sources, the system periodically pulls updated data via RESTful API or database connection. Ambient light data comes from the ambient light sensors on the street light nodes themselves, sampled every 1 minute; traffic flow data is obtained from the traffic management platform, updated every 5 minutes; personnel activity data is pushed through the video analysis system, updated every 10 minutes; weather visibility data is obtained from the meteorological service platform, updated every 30 minutes.

[0055] Data preprocessing and validation. The data interface layer performs preliminary processing on the received raw data. First, it performs data integrity validation, removing data packets missing key fields. Second, it performs reasonableness checks, marking or filtering outliers that exceed physical limits (such as duty cycle > 100% or negative power). Third, it performs time alignment, mapping data from different data sources and sampling frequencies to a unified time grid (e.g., using 1 minute as the base time unit). For sensor noise, it uses sliding window averaging or median filtering methods for smoothing. For data loss caused by communication delays or packet loss, the system uses linear interpolation or nearest neighbor interpolation methods to complete the data, and adds an "uncertainty flag" to the interpolated data for use by the subsequent uncertainty assessment module.

[0056] Data aggregation and caching. Preprocessed data is aggregated into a unified data storage layer, such as the time-series database InfluxDB or the relational database PostgreSQL. The system maintains two types of data views: a real-time data view, which stores high-frequency data from the most recent hour for the lighting status construction module to access in real time; and a historical data view, which stores long-term operational data for use by the model training and optimization module.

[0057] The lighting state construction module is responsible for mapping discrete street light node brightness parameters to a continuous spatial illuminance distribution, forming an overall lighting state model (light field model). This model serves as an intermediate state representation of the system, uniformly characterizing the lighting level at different spatial locations, and providing state input for subsequent situation analysis and dimming decisions.

[0058] Method for constructing the light field model. This embodiment adopts a road segmentation light field model. Specifically, a 2-kilometer road is divided longitudinally into 200 road segment units (each unit is 10 meters long) and laterally into 3 lane areas (left lane, middle lane, and right lane). Therefore, the overall space is discretized into 200 × 3 = 600 spatial grids. For each spatial grid (i, j), its illuminance value E(i, j, t) is composed of the sum of the illuminance contributions from several surrounding street light nodes. Assuming that the current brightness of street light node k is L_k(t) (proportional to the driving duty cycle), its contribution to the illuminance of grid (i, j) can be expressed as:

[0059] ;

[0060] Where G_k(i,j) is the light propagation function from streetlight k to grid (i,j), which comprehensively considers factors such as distance attenuation, luminaire light distribution curve, installation height, and angle. During system initialization, G_k(i,j) is calibrated using lighting simulation software or on-site measured data. The total illuminance of grid (i,j) is the sum of the contributions from all streetlights.

[0061] ;

[0062] E_ambient(t) represents the ambient light contribution, such as moonlight or roadside building lights, which is obtained by measuring the ambient light sensor and performing spatial interpolation.

[0063] The light field model is dynamically updated over time to reflect changes in lighting conditions. The system employs a strategy combining periodic updates and event-triggered updates. Periodic updates use a 1-minute timeframe. At the end of each period, the system reads the latest luminance values ​​L_k(t) of all streetlight nodes, recalculates the illuminance E(i,j,t) of each grid, and updates the overall light field model. Event-triggered updates are executed immediately upon detecting significant state changes, such as when a streetlight node malfunctions (luminance suddenly drops to zero) or when a dimming command causes a step change in brightness. The system immediately updates the light field model of the affected area without waiting for the next period.

[0064] To improve computational efficiency, the system employs an incremental update strategy: only the illuminance of streetlight nodes experiencing brightness changes and the grid illuminance within their affected range are recalculated; the illuminance values ​​of unaffected grids are directly reused from the previous cycle. The affected range is determined by a pre-calculated illuminance contribution threshold, for example, when... When lux, the effect of street light k on grid (i, j) is considered negligible.

[0065] The lighting situation analysis module infers and predicts changes in lighting demand based on the overall lighting state model and external state data. This embodiment employs a situation analysis method based on a structural causal model, explicitly modeling multiple factors influencing lighting demand and their causal relationships, thereby achieving a mechanistic understanding and forward-looking inference of demand changes.

[0066] The structural causal model is constructed by building a directed acyclic graph (DAG) to represent the causal relationships between various influencing factors. Nodes in the graph represent state variables, and edges represent causal paths. In this embodiment, the main nodes include: ambient brightness (X_ambient), traffic flow (X_traffic), pedestrian activity density (X_pedestrian), weather visibility (X_visibility), time period (X_time), current illuminance level (X_illuminance), and target illuminance requirement (Y_target). The causal relationships are set based on lighting engineering knowledge and historical data analysis. For example: ambient brightness → target illuminance requirement (the brighter the environment, the lower the demand for artificial lighting, a negative causal effect); traffic flow → target illuminance requirement (the greater the traffic flow, the higher the safety requirements for road lighting, a positive causal effect); weather visibility → target illuminance requirement (lower visibility, such as in smoggy weather, requires increased lighting compensation, a positive causal effect); time period → traffic flow and pedestrian activity density (increased traffic flow and pedestrian activity during evening rush hour, a positive causal effect).

[0067] Each causal edge is configured with two parameters: a causal influence coefficient α and a time delay τ. The causal influence coefficient represents the strength of the impact of changes in upstream nodes on downstream nodes, for example... This indicates that for every 10% increase in traffic flow, the target illuminance requirement increases by 3%. The time delay parameter represents the propagation lag of causal effects, for example... The minute parameter represents a 5-minute delay in adjusting lighting requirements following changes in weather visibility. These parameters are estimated using causal inference algorithms based on historical data (such as PC algorithms or constraint-based learning methods) and are updated periodically.

[0068] Based on the structural causal model, the system executes the following inference process: First, it reads the observation values ​​of each node at the current time (such as traffic flow X_traffic(t), ambient brightness X_ambient(t), etc.); second, it calculates the target illuminance demand Y_target(t) by propagating the influence step by step from upstream nodes to downstream nodes according to the causal relationship graph. The specific calculation adopts a linear weighted combination form.

[0069] ;

[0070] Where Y_base is the baseline illuminance requirement. Let X_i represent the causal coefficients of each influencing factor, X_i·(t-τ_i) represent the upstream node value after considering time delay, and X_i, baseline represent the corresponding baseline values.

[0071] For future demand forecasting, the system uses a time series forecasting model to predict the values ​​of each upstream node in future time periods. The predicted values ​​are then used to calculate future target illuminance requirements using a causal model. _target(t+Δt). For example, the system can predict the trend of lighting demand changes over the next 15 minutes, providing forward-looking information for dimming decisions. Furthermore, the system can perform counterfactual inference, answering the question, "How will lighting demand change if a certain factor changes?" For instance, if a large number of vehicles are predicted to enter the road segment within the next 10 minutes, i.e., traffic flow is expected to increase by 50%, the system can calculate the target illuminance demand increment under this scenario using a causal model, thereby preparing for dimming in advance.

[0072] In actual operation, the system faces multiple sources of uncertainty: sensor measurement errors, data loss due to communication delays, inaccuracy of external data, estimation errors of model parameters, and inherent uncertainties in future predictions. The role of the uncertainty assessment module is to quantify these uncertainties and pass the assessment results to the dimming decision module, enabling the decision to remain robust under uncertain conditions.

[0073] For both sensor data and external data, the system employs Bayesian inference to assess their uncertainties. Taking ambient light data as an example, assuming the sensor measurement is z_ambient and the actual ambient brightness is X_ambient, a measurement error exists between the two. The system maintains a prior distribution P(X_ambient) (based on historical data and time patterns), and combines it with the observed value z_ambient to obtain the posterior distribution P(X_ambient | z_ambient) through Bayesian updates:

[0074] ;

[0075] The variance of the posterior distribution characterizes the uncertainty of the data. If the sensor has not been updated for a long time or the data is completed through interpolation, the system will increase the corresponding uncertainty estimate. For missing data, the uncertainty is set as the variance of the prior distribution.

[0076] For both the light field model and the causal relationship model, the system assesses the uncertainties arising from parameter estimation errors and model structure errors. Taking the light field model as an example, the calibration of the illumination propagation function G_k(i,j) has errors. The system obtains the parameter distribution by fitting multiple measured data sets and calculates the confidence interval for the predicted illuminance. For the illuminance prediction E(i,j,t) at grid (i,j), the system provides not only the point estimate but also a 95% confidence interval [E_lower(i,j,t), E_upper(i,j,t)]. The width of the confidence interval reflects the level of uncertainty in the model. For the parameters of the causal relationship model (such as the causal influence coefficient α), the system also maintains their probability distribution. During situation inference, the system can propagate parameter uncertainty using a Monte Carlo sampling method: multiple samples are drawn from the parameter distribution, and the target illuminance demand is calculated for each sample to obtain the demand prediction distribution P(Y_target). The variance of this distribution characterizes the uncertainty of the situation analysis results.

[0077] Finally, the uncertainty assessment module outputs the following results: (1) confidence parameter, which represents the credibility of the data or prediction results, with a value range of 0-1; (2) uncertainty measure, such as standard deviation or variance; (3) probability distribution, such as posterior distribution P(X | observed data) or prediction distribution P(Y_target | historical data).

[0078] The dimming decision and centralized control module integrates lighting situation analysis results and uncertainty assessment results to generate and implement collaborative dimming control strategies for multiple smart street light nodes. The core task of this module is to determine the optimal brightness adjustment scheme for each street light node while meeting lighting target constraints and energy consumption constraints.

[0079] The dimming decision objectives of the system include two aspects: first, to ensure that the lighting level of the entire road segment meets the requirements, that is, the actual illuminance E(i,j,t) of each grid is as close as possible to the target illuminance Y_target(i,j,t), while avoiding local areas being too dark or too bright; second, to minimize the total energy consumption while meeting the lighting requirements. Therefore, the decision problem can be formalized as a multi-objective optimization problem:

[0080] ;

[0081] The constraints include: the brightness L_k of each street light is within an adjustable range (e.g., 30%-100%), the illuminance of each grid meets the minimum safety standard (e.g., E(i,j,t) ≥ E_min), and the dimming rate is limited (to avoid frequent flickering). Where P_k(L_k) is the power consumption of street light k at brightness L_k. This is the trade-off coefficient between lighting quality and energy saving.

[0082] In situations involving uncertainty, the system employs robust optimization or risk-averse decision-making strategies. Specifically, when the uncertainty assessment module provides a high uncertainty metric, the system tends to adopt a conservative dimming strategy: for example, if the target illuminance demand Y_target has significant uncertainty (high prediction variance), the system will reserve a certain illuminance margin during dimming to avoid insufficient lighting due to underestimation of demand. In practice, the system replaces the target illuminance Y_target with its upper quantile (e.g., the 90th quantile) to ensure that lighting demand is met in most possible scenarios. Conversely, when uncertainty is low, the system can adopt a more aggressive energy-saving strategy, adjusting the brightness to a level closer to the predicted demand. Furthermore, for situations with long prediction time spans and high uncertainty, the system employs a gradual dimming strategy: first, small-amplitude dimming is performed, feedback is observed, and then adjustments are made gradually to avoid discovering a large deviation in demand prediction after a one-time large-amplitude dimming.

[0083] This embodiment employs a multi-agent collaborative mechanism to achieve distributed decision-making. The system abstracts each streetlight node as a software agent. Each agent shares the overall lighting state model (light field model), but each is responsible for optimizing its local objective. The local objective of each agent i is to minimize the weighted sum of the illuminance deviation within its jurisdiction and its own energy consumption:

[0084] ;

[0085] Where R_i is the set of grids influenced by agent i, and β is the energy consumption weight. Agents negotiate through message passing: each agent calculates its optimal brightness adjustment based on the current light field model and the brightness plans of neighboring agents; then it broadcasts the adjustment plan to neighboring agents; upon receiving the message, neighboring agents update their light field models and recalculate their optimal policies. This process iterates until all agents' policies converge (the adjustment amount is less than a threshold for two consecutive rounds) or the maximum number of iterations is reached. After convergence, each agent sends the final brightness adjustment command to the corresponding street light node for execution.

[0086] After the decision is made, the dimming control module generates specific dimming commands. These commands include the NodeID of the target streetlight node, the target duty cycle (target brightness percentage), the dimming rate (e.g., changes per second), and the execution time. To avoid impacting the power grid from simultaneous large-scale dimming, the system schedules the dimming commands in a timing sequence: the 120 streetlights are divided into several batches, with each batch dimming sequentially at several-second intervals, ensuring a smooth change in grid load. The dimming commands are sent to each streetlight node via a wireless network. Upon receiving the commands, the local control unit of each streetlight node smoothly adjusts the LED driver duty cycle at the specified rate, achieving a gradual transition in brightness.

[0087] After the dimming command is executed, the feedback and optimization module is responsible for collecting operational feedback information and updating or adjusting the system's state model or dimming control strategy based on the feedback, forming closed-loop control and continuous optimization.

[0088] After dimming is executed, the system continuously monitors the following feedback data: (1) the actual brightness output of the street light nodes, obtained by reading the drive duty cycle or the luminous flux output of the light source; (2) the actual illuminance at each grid location, obtained by using a road illuminance sensor (if deployed) or calculated based on the updated light field model; (3) energy consumption data, recorded by the power metering unit before and after dimming; and (4) subsequent changes in external state data, such as the actual evolution of traffic flow or personnel activity after dimming. The data collection cycle of the feedback data is synchronized with the dimming execution cycle, usually 1-5 minutes.

[0089] The system compares the feedback data with the expected results, identifies model errors, and updates the model parameters. Taking the light field model as an example, if there is a persistent deviation between the actual illuminance E_actual(i,j) of a certain grid and the model prediction E_predicted(i,j), the system infers that the illumination propagation function G_k(i,j) may have a calibration error, and then uses gradient descent or Kalman filtering methods to correct G_k(i,j).

[0090] The system also performs Bayesian updates based on feedback: it uses observed changes in demand as new evidence, combines them with the prior distribution to obtain the posterior distribution of the parameters, thereby gradually improving the model's prediction accuracy.

[0091] In addition to updating model parameters, the system can also employ reinforcement learning methods to optimize the dimming strategy itself. Specifically, the system models the dimming decision-making process as a Markov Decision Process (MDP): the state includes the current light field model, external state data, and uncertainty assessment results; the action is the brightness adjustment scheme for each street light; and the reward is a comprehensive score of lighting quality and energy-saving effect (a negative objective function value). The system uses deep reinforcement learning algorithms (such as DQN or PPO) to train the policy network, continuously optimizing the strategy through interaction with the environment (dimming-feedback-update). Training data comes from historical operation records and real-time feedback, and the parameters of the policy network are updated periodically. After long-term training, the system can learn to automatically select the optimal dimming strategy in different scenarios (such as sunny nights, rainy / foggy weather, peak hours, etc.) without the need for manually setting complex rules.

[0092] The feedback and optimization module is also responsible for detecting system malfunctions. If the actual brightness reported by a street light node deviates significantly from the target brightness, the system determines that the node may be malfunctioning (e.g., driver failure or communication failure), immediately triggers an alarm, and marks the node as unavailable. The system then recalculates the brightness of neighboring street lights to compensate for the lighting loss caused by the malfunctioning node. If the actual illuminance of multiple grids remains below the safety standard, the system determines that there may be widespread lighting insufficiency and triggers an emergency dimming mode: forcing all available street lights to increase to maximum brightness to ensure basic safety, while simultaneously sending a warning message to maintenance personnel.

[0093] Through the aforementioned feedback and optimization mechanisms, the system achieves adaptive model evolution, continuous strategy improvement, and rapid response to abnormal situations, forming a closed-loop control system with self-learning capabilities. In long-term operation, the system's lighting control accuracy, energy-saving effect, and robustness are continuously improved.

[0094] Example 2

[0095] This embodiment, based on Embodiment 1, further details the specific calculation processes and algorithm implementations of each core module of the system. The purpose of this embodiment is to enable those skilled in the art to more clearly understand and implement the construction method of the overall lighting state model, the inference mechanism of causal relationship analysis, the quantitative assessment method of uncertainty, and the solution process of multi-node collaborative optimization by elaborating on the detailed calculation steps of key technical aspects, providing specific numerical examples and algorithm flows. This embodiment uses the same system deployment scenario as Embodiment 1, focusing on the implementation details at the computational level.

[0096] The light field model maps the brightness output of discrete street light nodes to a spatially continuous illuminance distribution. The following details the modeling method for the light propagation function, the calculation steps for grid illuminance, and the specific process for model updates.

[0097] For a streetlight node k, its illumination contribution to the spatial grid (i, j) depends on the distance, angle, and light distribution characteristics of the luminaire. Using a simplified inverse squared attenuation model combined with a light distribution correction coefficient, the illumination propagation function G_k(i, j) is expressed as:

[0098] ;

[0099] Where I_0 is the nominal luminous intensity of the street lamp (in cd), θ_k(i,j) is the angle between the ray from street lamp k to grid (i,j) and the downward direction perpendicular to the street lamp, and d_k(i,j) is the spatial distance from street lamp k to grid (i,j). The comprehensive efficiency coefficient takes into account factors such as luminaire efficiency and atmospheric transmittance, with a typical value of 0.7-0.9. The method approximates the directional characteristics of the street light distribution curve.

[0100] In the specific calculation, the spatial distance is first calculated based on the spatial coordinates (x_k, y_k, h_k) of street lamp k and the center coordinates (x_ij, y_ij, 0) of the grid (i, j) (with the road surface height set to 0).

[0101] ;

[0102] The cosine of the included angle is calculated as follows:

[0103] ;

[0104] In this embodiment, streetlight k is located at coordinates (0, 0, 8), with an installation height of 8 meters. Grid (i, j) is located at (5, 3, 0), approximately 5.83 meters horizontally from the streetlight. Therefore:

[0105] · rice;

[0106] · ;

[0107] · ;

[0108] If I_0 = 10000 cd and η = 0.8, then: G_k(i,j) = (10000 × 0.527) / 98 × 0.8≈ 43.0 (cd / m² needs to be multiplied by the solid angle factor to convert to lux, which is simplified to the direct illuminance contribution coefficient here);

[0109] For grid (i, j), its total illuminance E(i, j, t) is the sum of contributions from all streetlights and ambient light. In a scenario with 120 streetlights, only streetlights closer to the grid (e.g., less than 30 meters) are considered in the calculation to improve efficiency. Let the current duty cycle of streetlight k be D_k(t) (range 0-1), and its luminance output L_k(t) be linearly related to the duty cycle:

[0110] ;

[0111] Where L_max is the luminous flux output of the streetlight at full power. The grid illuminance is calculated as follows:

[0112] ;

[0113] Where N(i,j) is the set of streetlights that make a significant contribution to grid (i,j), and E_ambient(i,j,t) is the ambient illuminance, which is obtained by spatial interpolation of the measurements from nearby sensors.

[0114] Consider a grid (100, 2) where four streetlights significantly contribute to its size, numbered k=50, 51, 52, and 53. The parameters and contributions of each streetlight are calculated in Table 1 below.

[0115] Table 1: Contribution Calculation Table Streetlight K D_k(t) G_k(100, 2) Illuminance contribution (lux) 50 0.60 35.2 0.60×35.2=21.12 51 0.65 42.8 0.65×42.8=27.82 52 0.60 38.5 0.60×38.5=23.10 53 0.70 30.1 0.70×30.1=21.07

[0116] Assuming the ambient light E_ambient(100, 2, t) = 2.5 lux (including contributions from moonlight and surrounding building light), the total grid illuminance is: E(100, 2, t) = 21.12 + 27.82 + 23.10 + 21.07 + 2.5 = 95.61 lux;

[0117] Incremental update of the light field model. When the duty cycle of street lamp k changes from D_k(t-1) to D_k(t), only the affected mesh set R_k needs to be updated. For mesh (i, j) ∈ R_k, the illuminance update is:

[0118] ;

[0119] In this embodiment, the duty cycle of street light 51 is adjusted from 0.65 to 0.80, so the illuminance of grid (100, 2) is updated as follows: E(100, 2, t) = 95.61 + L_max × (0.80 - 0.65) × 42.8 = 95.61 + 6.42 = 102.03 lux;

[0120] By using incremental updates, the system avoids full recalculation of all 600 grids, significantly improving real-time performance.

[0121] Causal relationship models are used to analyze the triggering mechanisms and impact paths of changes in lighting demand. The following details the specific steps involved in constructing the causal graph, estimating parameters, and performing inference calculations.

[0122] The structure of the causal graph is defined as follows. The causal directed acyclic graph constructed in this embodiment contains 7 nodes and 9 edges. The node definitions are as follows: X1: Time period (value: 0-23 represents hours) •X2: Ambient brightness (unit: lux, sensor measurement) • X3: Weather visibility (in km) •X4: Traffic flow (vehicles per hour) •X5: Personnel activity density (per person / area) • X6: Current illuminance level (in lux) • Y: Target illuminance requirement (in lux).

[0123] Causal edges and their semantics are as follows: • X1 → X4: Time period affects traffic flow (increased traffic during evening rush hour) • X1 → X5: Time period affects personnel activities (reduced nighttime activities) • X2 → Y: Ambient brightness affects lighting demand (negative correlation) • X3 → Y: Visibility affects lighting requirements (low visibility requires compensation) • X4 → Y: Traffic flow affects lighting requirements (higher traffic flow requires higher illuminance) • X5 → Y: Human activity affects lighting requirements (higher activity requires higher illuminance) ·X6 → Y: Current illuminance affects adjustment needs (closed-loop feedback).

[0124] Estimation of the causal influence coefficient. For the edge X_i → X_j, the causal influence coefficient α_{i→j} is estimated through regression analysis of historical data.

[0125] Using historical data from the past 30 days, with one sample per hour for a total of 720 samples, the parameters were estimated using the least squares method.

[0126] In this embodiment, for edge X4 → Y (traffic flow → target illuminance), regression analysis yields the following:

[0127] α_{4→Y} = 0.012, meaning that for every increase of 1 vehicle / hour in traffic flow, the target illuminance requirement increases by 0.012 lux;

[0128] • When the time delay τ_{4→Y} = 0, changes in traffic immediately affect demand;

[0129] • Confidence interval: [0.010, 0.014];

[0130] Similarly, the estimation results for the other edges are as follows:

[0131] • α_{2→Y} = -0.15, meaning that for every 1 lux increase in ambient brightness, the demand decreases by 0.15 lux;

[0132] • α_{3→Y} = -2.5, meaning that for every 1 km decrease in visibility, the demand increases by 2.5 lux;

[0133] ·α_{5→Y} = 0.08, for every increase of 1 person in personnel density, the demand increases by 0.08 lux.

[0134] The calculation process for situational inference is as follows: Given the observation values ​​at the current time t, the system calculates the values ​​of each node according to the topological sorting order. The specific steps are as follows:

[0135] Step 1: Read exogenous variables, variables without parent nodes: ·X1(t) = 19 (current time 19) ·X2(t) = 5.2 lux (measured by ambient light sensor) ·X3(t) = 8.0 km (meteorological platform data).

[0136] Step 2: Calculate intermediate nodes, variables with parent nodes:

[0137] X4(t) = β_{0,4} + α_{1→4} · X1(t) + ε_4; β_{0,4} = 200, α_{1→4} = 50, which are the evening peak coefficients, then: X4(t) = 200 + 50 × (19 - 18) = 250, the current period is the evening peak; the actual real-time observation value of the traffic management system is: X4(t) = 320;

[0138] Similarly, X5(t) can be obtained from the observation data: X5(t) = 45 people;

[0139] Step 3: Calculate the target illuminance requirement Y(t). Baseline value Y_base = 30 lux (nighttime road standard), and sum the causal effects of various factors:

[0140] Y(t) = Y_base + α_{2→Y} · [X2(t) - X2_baseline] + α_{3→Y} · [X3(t) - X3_baseline] + α_{4→Y} · [X4(t) - X4_baseline] + α_{5→Y} · [X5(t) -X5_baseline];

[0141] The baseline values ​​are set as follows: nighttime ambient light baseline: X2_baseline = 0, normal visibility: X3_baseline = 10 km, X4_baseline = 200 vehicles / hour, X5_baseline = 30 people.

[0142] Substituting the values: Y(t) = 30 + (-0.15)×(5.2 - 0) + (-2.5)×(8.0 - 10) + 0.012×(320 - 200) + 0.08×(45 - 30) = 30 - 0.78 + 5.0 + 1.44 + 1.2 = 36.86 lux;

[0143] Therefore, the target illuminance requirement at the current moment is 36.86 lux.

[0144] For a future time t+Δt, the system first predicts changes in upstream variables. Traffic flow is predicted using the ARIMA model:

[0145] In this embodiment, the ARIMA(2,0,1) model predicts X4(t+15min) = 340 vehicles / hour. With other variables remaining constant, the future demand is:

[0146] Y(t+15min) = 30 - 0.78 + 5.0 + 0.012×(340-200) + 1.2 = 37.10 lux;

[0147] This prediction is used for dimming decisions. The uncertainty assessment module quantifies the reliability of the data, model, and predictions. The specific computational steps for Bayesian inference and multi-model ensemble are detailed below.

[0148] Taking ambient light sensor data as an example, assuming the sensor measurement z_ambient = 5.2 lux, the actual ambient brightness X2 follows a normal distribution, and the measurement error σ_sensor = 0.5 lux. The prior distribution is set based on historical patterns. , where μ_prior is set to 5.0 lux based on the historical average of the current time period (19:00), and σ_prior = 1.0 lux.

[0149] Bayesian update calculates the posterior distribution: likelihood function: Posterior distribution ;

[0150] Formulas for calculating the posterior mean and variance:

[0151] ;

[0152] Substituting the values, we get μ_post = (5.0 / 1.0 + 5.2 / 0.25) / (1 / 1.0 + 1 / 0.25) = (5.0 +20.8) / (1 + 4) = 25.8 / 5 = 5.16 lux;

[0153] Therefore, the posterior estimate of ambient brightness is 5.16 lux, with an uncertainty (standard deviation) of 0.447 lux. Compared to the prior uncertainty of 1.0 lux, the observed data significantly reduce the uncertainty.

[0154] The estimated causal influence coefficient α_{4→Y} is 0.012, with a 95% confidence interval of [0.010, 0.014], which can be approximated as a normal distribution. When the traffic flow is X4 = 320, the contribution of this factor to the target illuminance is:

[0155] ;

[0156] That is, the mean of the contribution is 1.44 lux and the standard deviation is 0.12 lux.

[0157] Considering the uncertainties of all factors, the total uncertainty of the target illuminance Y(t) is calculated using the error propagation formula. Assuming the factors are independent, then:

[0158] ;

[0159] Substituting the terms, the standard deviations of the other terms in this embodiment are 0.08, 0.15, and 0.10 lux, respectively. ;

[0160] Therefore, the target illuminance requirement Y(t) = 36.86 ± 0.21 lux, which is 68% confidence level.

[0161] For traffic flow prediction in the next 15 minutes, the system was tested with other models: ARIMA(2,0,1): predicted X4(t+15min) = 340 vehicles / hour; Exponential smoothing: predicted X4(t+15min) = 335 vehicles / hour; LSTM neural network: predicted X4(t+15min) = 348 vehicles / hour.

[0162] Calculate the mean and standard deviation of the prediction results: Mean = (340 + 335 + 348) / 3 = 341 vehicles / hour;

[0163] Standard deviation = Vehicles per hour.

[0164] The large degree of divergence in the forecasts (standard deviation of approximately 1.6% of the mean) indicates a high degree of uncertainty regarding future traffic flow.

[0165] Propagating the uncertainty of traffic flow forecasting to illumination demand: σ_{Y,future} = α_{4→Y} · σ_{X4,future} = 0.012 × 5.35 = 0.064 lux.

[0166] Considering both the uncertainty of the model parameters and the uncertainty of the forecast, the total uncertainty of future illuminance demand is: ;

[0167] The uncertainty assessment module outputs the following results for dimming decision-making:

[0168] Confidence parameter: confidence = 1 / (1 + σ_Y, total) = 1 / (1 + 0.22) = 0.82 (normalized to 0-1);

[0169] Uncertainty measure: uncertainty = σ_Y, total = 0.22 lux;

[0170] Probability distribution: P(Y) ~ N(36.86, 0.22²), which can provide illuminance intervals at different confidence levels, such as the 90% confidence interval [36.86 - 1.645×0.22, 36.86 + 1.645×0.22] = [36.50, 37.22] lux.

[0171] The dimming decision module solves for the optimal brightness adjustment scheme for multiple nodes based on the overall light field model, target illuminance requirements, and uncertainty assessment results.

[0172] ;

[0173] Constraints: For all k, D_min ≤ D_k ≤ D_max; For all i, j, E(i, j) ≥ E_min; |D_k(t) - D_k(t-1)| ≤ ΔD_max (dimming rate limit to avoid frequent flickering);

[0174] Where P_k(D_k) is the power consumption of street light k (approximately linear relationship P_k = P_max · D_k), λ is the lighting quality weight coefficient (typical value 1000-10000), and w_{ij} is the grid weight (high priority areas such as pedestrian crossings have higher weight).

[0175] The relationship between illuminance E(i,j) and duty cycle D_k is established through the light field model: E(i,j) = Σ_k L_max ·D_k · G_k(i,j) + E_ambient(i,j);

[0176] Robust optimization considering uncertainty. When uncertainty is high, the system employs a robust optimization strategy, replacing the target illuminance Y_target with its upper quantile Y_robust:

[0177]

[0178] Where κ is the risk coefficient (typical value 1.0-1.5), and σ_Y is the standard deviation of illuminance requirement. In the aforementioned case, if Therefore: Y_robust = 36.86 + 1.2 × 0.22 = 37.12 lux;

[0179] Ensure that lighting needs are met in most uncertain scenarios.

[0180] The Alternating Direction Multiplier Method (ADMM) is used for distributed solution. The 120 streetlights are divided into 10 regional groups of 12 lights each. Each group optimizes the local objective in parallel, and global consistency is coordinated by Lagrange multipliers.

[0181] The algorithm flow is as follows:

[0182] Initialization: Set all (Using the duty cycle of the previous cycle), iteration number iter = 0;

[0183] Iterate until convergence or the maximum number of iterations is reached:

[0184] Each region group g solves local subproblems in parallel. For region group g, the duty cycle of the streetlights under its jurisdiction {D_k |k∈G_g} is optimized. This subproblem is a convex quadratic programming problem, which is solved using the gradient descent method. After updating, it is projected onto the constraint range [D_min, D_max].

[0185] Global synchronization: The central node collects the update results of each region group, calculates the global average, and checks the constraints.

[0186] Example 3

[0187] This embodiment focuses on the specific implementation of the multi-agent collaboration and game theory mechanism in the dimming decision and centralized control module, as well as the detailed process of the feedback and optimization module adaptively updating each component of the system. This embodiment adopts the same system deployment scenario as Embodiment 1.

[0188] The types and responsibilities of intelligent agents include three categories of functional software intelligent agents:

[0189] Node Agents: Each street light node corresponds to one node agent, totaling 120. Each node agent is responsible for optimizing the brightness adjustment scheme of its own node. Each node agent has its own optimization goal: to minimize its own energy consumption while meeting the lighting needs of its affected area. Node agents can access a shared overall lighting state model to understand the illuminance distribution across the entire road segment and can perceive the decision-making intentions of neighboring node agents.

[0190] Regional Agent: The 120 streetlight nodes are spatially divided into 10 regions (each region is approximately 200 meters long and contains 12 nodes), with one regional agent configured in each region. The regional agent is responsible for coordinating the decisions of all node agents within its region, ensuring the uniformity and continuity of lighting distribution and preventing abrupt changes in brightness caused by excessive differences between adjacent streetlights. The regional agent is also responsible for coordinating the region boundary with adjacent regions.

[0191] Global Coordinating Agent: At the system level, a global coordinating agent is configured to supervise the decisions of all regional agents, ensuring that the lighting target constraints and energy consumption constraints for the entire road segment are met. The global coordinating agent does not directly control streetlight brightness; instead, it guides the decision-making direction of agents at all levels by adjusting constraint parameters or issuing coordination commands.

[0192] The system incorporates both lighting target constraints and energy consumption constraints when making dimming decisions, and achieves multi-objective balance through the setting of constraints and optimization objectives.

[0193] Lighting target constraints include: Minimum illuminance safety constraint: The illuminance at each grid location must not be lower than the safety standard, i.e., E(i,j) ≥ 20 lux, to ensure basic vehicle and pedestrian safety. Target illuminance achievement constraint: The illuminance in key areas (such as pedestrian crossings and intersections) should be as close as possible to the target value, with a deviation not exceeding ±15%. Illuminance uniformity constraint: The illuminance ratio between adjacent grids should be within a reasonable range, such as 0.4 ≤ E(i,j) / E(i+1,j) ≤ 2.5, to avoid abrupt changes in brightness.

[0194] Energy consumption constraints include: total power limit constraint: the total power of all streetlights along the entire road section does not exceed the set limit; single-lamp dimming range constraint: the duty cycle of each streetlight is limited to the range of 30%-100%, ensuring energy saving space while avoiding excessively low brightness that would shorten the lifespan of the lamps; dimming change rate constraint: the duty cycle change of a single dimming operation does not exceed 20%, and the dimming frequency does not exceed once every 10 minutes, avoiding frequent switching that could affect equipment lifespan; the system incorporates both lighting quality and energy consumption into the optimization objectives through a weighted approach. The global coordinating agent dynamically adjusts the weight coefficients according to the current scenario: increasing the weight of lighting quality during evening peak hours to prioritize safety; and increasing the weight of energy saving during low-traffic periods at night to prioritize reducing energy consumption.

[0195] The multi-agent collaborative mechanism employs a consensus-based iterative negotiation algorithm. Through information exchange and policy adjustments, each agent gradually converges to a globally consistent dimming scheme.

[0196] The collaboration process is as follows:

[0197] Initialization Phase: Based on the current lighting situation analysis and uncertainty assessment results, the global coordinating agent determines the overall objective of this round of dimming (such as the average target illuminance across the entire road segment and the total energy consumption budget) and broadcasts it to all regional agents. Each regional agent decomposes the objective into sub-objectives for its region and communicates them to its respective node agents.

[0198] Local optimization phase: Each node agent calculates its initial dimming scheme based on the shared overall lighting state model. Node agent i reads the illuminance distribution of its affected area in the current light field model, identifies grids with insufficient or excessive illuminance, and calculates the required brightness adjustment. During the calculation, the node agent assumes that the brightness of other nodes remains unchanged (or uses the decision value from the previous round), and optimizes its own duty cycle D_i accordingly.

[0199] Information exchange phase: Each node agent reports its preliminary dimming plan to its corresponding regional agent. The regional agent collects the plans from all nodes within its region and checks whether they meet the lighting uniformity constraints within the region. If it finds that brightness adjustments by adjacent nodes would cause a sudden change in illuminance, the regional agent sends a coordination request to the relevant node agents, suggesting an adjustment to the duty cycle.

[0200] Negotiation and Adjustment Phase: Upon receiving the coordination request, the node agent recalculates the dimming scheme, seeking a balance between its own optimization goals and regional coordination requirements. For example, a node agent originally planned to reduce the duty cycle from 60% to 40% to save energy, but a regional agent pointed out that this would lead to excessive brightness differences with neighboring nodes (70% duty cycle). The node agent then re-optimizes, changing the adjustment target to 50%, achieving a compromise between energy saving and uniformity.

[0201] Global Verification Phase: Each regional agent summarizes its local dimming scheme and reports it to the global coordination agent. The global coordination agent calculates the total illuminance distribution and total energy consumption across the entire road segment based on the updated light field model. If any global constraints are found to be unmet, the global coordination agent sends adjustment instructions to the relevant regional agents, requesting further optimization.

[0202] Iterative convergence: The processes of local optimization, information exchange, negotiation adjustment, and global verification described above are performed iteratively. The system's convergence condition is set as follows: the duty cycle change of all nodes is less than 1% in two consecutive iterations, or the number of iterations reaches the upper limit (e.g., 15 times). In actual operation, the system usually converges after 5-8 iterations.

[0203] In this embodiment, the coordination process of three adjacent streetlights (nodes 50, 51, and 52) within a certain area is considered. Initially, the duty cycle of all three is 60%. After the first round of optimization, node 50 is planned to decrease to 45% due to excessive illuminance in its affected area; node 51 is planned to maintain 60%; and node 52 is planned to increase to 75% due to increased demand in its affected area. The regional agent detects significant brightness differences between nodes 50 and 51, and between nodes 51 and 52, and sends coordination suggestions to nodes 50 and 52. After the second round of optimization, node 50 is adjusted to 50%, and node 52 is adjusted to 70%, with the brightness gradient becoming more gradual. After the third iteration, the scheme converges to: node 50 with a duty cycle of 52%, node 51 with a duty cycle of 58%, and node 52 with a duty cycle of 68%, satisfying the lighting needs of their respective areas while maintaining good uniformity.

[0204] In the implementation of multi-agent game theory, in some cases, the optimization objectives of different agents may conflict, and the system uses game theory to make decisions.

[0205] There is an overlapping lighting zone at the boundary of two adjacent regions. An agent in region A wants to reduce the brightness of the boundary streetlights to save energy, while an agent in region B wants to increase the brightness of the boundary streetlights to improve the lighting uniformity of its region. This creates a non-cooperative game situation. The dimming decision is modeled as a multi-player non-cooperative game. Each regional agent is a player in the game, and its policy space is the combination of the duty cycles of all nodes in its region. The payoff function of each player is the overall utility of its region (lighting quality minus energy cost). Given the policies of other players, each player optimizes its own policy to maximize its payoff.

[0206] The system employs an iterative optimal response algorithm to solve for the Nash equilibrium. In each iteration, each regional agent assumes that the policies of other regions are fixed and solves for its own optimal response policy. Specifically, agent A in region A calculates the optimal dimming scheme for its region based on the current brightness of the boundary streetlights in region B; simultaneously, agent B in region B performs the same calculation based on the policy of region A. After multiple iterations, the policies of each region gradually converge to the Nash equilibrium point, meaning that no single regional agent can gain a higher benefit by unilaterally changing its policy.

[0207] In this embodiment, there are two streetlights (node ​​60 belongs to region A, and node 61 belongs to region B) at the boundary between region A and region B, and the two lights are 15 meters apart. Initially, node 60 has a duty cycle of 55%, and node 61 has a duty cycle of 65%. Region A wants to reduce the duty cycle of node 60 to 45% to save energy, but this will lead to insufficient illumination in the boundary area, affecting region B. The game iteration process is as follows:

[0208] Round 1: In Region A, node 60 was reduced to 45%, while in Region B, to compensate for the loss of illuminance, node 61 was increased to 75%.

[0209] Round 2: In Region A, after node 61 was found to have increased, the boundary illumination was found to be sufficient, so node 60 was further reduced to 40%. In Region B, node 61 was adjusted accordingly to 72%.

[0210] Round 3: The policy change is less than the threshold and converges to Nash equilibrium: Node 60 has a duty cycle of 42% and Node 61 has a duty cycle of 73%.

[0211] Under the balanced approach, region A achieved its energy-saving target (node ​​60 reduced from 55% to 42%), while region B maintained its boundary lighting quality by moderately increasing the brightness of node 61 (from 65% to 73%). Neither region could achieve a better result by unilaterally changing its strategy.

[0212] The system dynamically selects between cooperative or game-theoretic modes based on the consistency of objectives among agents. When the optimization directions of each agent are largely consistent (e.g., all aiming to reduce brightness and save energy), a cooperative mechanism is used to accelerate convergence; when there is a clear conflict of interest (e.g., a trade-off between energy saving and lighting quality), the system switches to game-theoretic mode to find an equilibrium solution. The global coordinating agent is responsible for monitoring and mode switching.

[0213] The feedback and optimization module collects operational feedback information after dimming is executed, and adaptively updates the causal relationship model, uncertainty assessment results, and dimming control strategy based on the feedback, enabling the system to have continuous learning and self-optimization capabilities.

[0214] After dimming is executed, the system continuously collects the following three types of feedback information:

[0215] Direct feedback information includes the actual brightness output of each street light node after dimming (reading the drive duty cycle), actual power consumption (measured by the power metering unit), and operating parameters such as light source temperature. Additionally, if illuminance sensors are deployed along the road section, the system can also obtain actual illuminance measurements at key locations. This direct feedback information is used to verify the accuracy of the light field model and the effectiveness of the dimming commands.

[0216] Indirect feedback information includes the actual evolution of external conditions over a period of time (e.g., 10-30 minutes) after dimming. For example, the system predicts that traffic flow will increase to 340 vehicles per hour in the next 15 minutes and adjusts the lighting accordingly; however, the actual traffic flow collected during operation is 355 vehicles per hour. This deviation between the predicted and actual values ​​constitutes indirect feedback, which is used to correct the parameters of the causal relationship model and the prediction model.

[0217] Effect Feedback Information: The overall effectiveness of dimming decisions is quantified through comprehensive evaluation indicators. The system calculates the following indicators: illumination compliance rate (the proportion of grid cells that meet the target illuminance range), energy saving rate (the percentage of energy saved compared to the baseline solution), illuminance uniformity, dimming response time, etc. This effect feedback information is used to evaluate the merits of the dimming control strategy and guide the adjustment of strategy parameters.

[0218] The parameters of the causal relationship model (causal influence coefficient and time delay) are continuously corrected through feedback data to improve the accuracy of situational analysis. The system has two types of update trigger conditions: First, cumulative triggering: after accumulating a certain number of feedback samples, batch parameter updates are performed. Second, deviation triggering: when the deviation between the model's prediction and the actual observed values ​​continuously exceeds a threshold, an emergency update is immediately performed.

[0219] An online learning algorithm is used to update the causal influence coefficient. Taking the influence of traffic flow on target illumination as an example, the original parameter is α_{traffic→target} = 0.012. The system records historical data for the past 30 days: the correspondence between the change in traffic flow ΔX_traffic and the change in target illumination ΔY_target. The causal coefficient is re-estimated through rolling window regression analysis. If the recent data trend indicates that the influence of traffic flow is increasing (possibly due to seasonal changes, road condition changes, etc.), the system updates the parameter to α_{traffic→target} = 0.014.

[0220] In this embodiment, the system predicted an increase in illuminance demand of 0.72 lux as traffic flow increased from 300 to 360 vehicles per hour (ΔX=60) based on α=0.012. The actual observed illuminance demand increased by 0.95 lux, with a deviation of 0.23 lux. Similar deviations were observed for five consecutive days. The system then initiated a parameter update, re-regressing based on the most recent 100 samples to obtain a new causality coefficient α=0.015. After the update, the same traffic flow change resulted in a predicted increase in illuminance demand of 0.90 lux, which is closer to the actual value.

[0221] The system identifies the optimal time delay for causal effects through cross-correlation analysis. In addition to parameter updates, the system can adjust the structure of the causal graph based on long-term feedback data. If a previously unmodeled factor (such as air quality index) is found to have a significant correlation with lighting demand, the system can add new nodes and edges to the causal graph. If a causal edge has no significant impact over a long period, it is removed from the model.

[0222] The confidence parameters and uncertainty measurement methods of the uncertainty assessment module are adjusted based on the accuracy of the feedback data.

[0223] In this embodiment, the system uses three models to predict future illuminance demand, with predicted values ​​of 36, 37, and 35 lux, respectively, and a standard deviation of 1.0 lux. The uncertainty metric is set to 1.0. The actual observed illuminance demand is 40 lux, a deviation of 4 lux, far exceeding the predicted uncertainty range. Continuous observation revealed that the standard deviation of the actual deviation is approximately 2.5 lux. The system updates the uncertainty correction factor from 1.0 to 2.5 / 1.0 = 2.5, and subsequently uses this factor to amplify the model divergence, making the uncertainty assessment more conservative yet more accurate.

[0224] The system sets thresholds for anomaly detection. If the false alarm rate is too high, the system relaxes the threshold to 3.5 times the standard deviation; if the false negative rate is too high, the system tightens the threshold to 2.5 times the standard deviation. Through feedback adjustments, the system gradually finds the optimal anomaly detection sensitivity.

[0225] The parameters of the dimming control strategy are optimized based on feedback. This includes: Adaptive adjustment of weighting coefficients: The balance coefficients of lighting quality weight λ and energy consumption weight in the dimming objective function are dynamically adjusted based on historical performance feedback, and self-learning of constraint thresholds: Constraint thresholds such as minimum illuminance safety standards and illuminance uniformity requirements can be fine-tuned based on actual operation feedback.

[0226] If no safety issues or user complaints are found in certain road sections where the illuminance is slightly below the standard value, the system can appropriately relax the minimum illuminance constraint for those road sections to increase energy-saving potential. Conversely, if feedback on insufficient lighting is received for certain road sections after meeting the standard value, the system will raise the illuminance target for those road sections.

[0227] In this embodiment, the initial minimum illuminance for a certain road section was set at 20 lux. After three months of operation, system analysis and feedback data revealed that the average nighttime traffic flow on this road section was only 50 vehicles per hour (far lower than the design value of 200 vehicles per hour), and there were no safety accidents or complaints. The system determined that the illuminance standard for this road section could be appropriately reduced. Through gradual adjustments—first reducing it to 18 lux and observing feedback for one month, then reducing it to 16 lux—the system ultimately stabilized the minimum illuminance standard for this road section at 17 lux, achieving an additional 8% energy saving while ensuring safety.

[0228] Optimization of agent collaboration strategy: The parameters during agent negotiation are adjusted based on historical convergence speed and decision quality. If convergence is not achieved after 15 iterations in some scenarios, the system increases the iteration limit to 20 iterations; if convergence is achieved in 5 iterations in most scenarios, the system lowers the limit to 10 iterations to save computation time.

[0229] Reinforcement Learning Strategy Optimization: The system employs deep reinforcement learning to continuously optimize the dimming strategy. Each dimming decision and its subsequent effect constitute a training sample (state, action, reward). The system stores these samples in an experience replay buffer and performs periodic batch training to update the policy network parameters. After long-term training, the policy network learns to automatically select the optimal dimming scheme in different scenarios without requiring manual rule adjustments.

[0230] The causal relationship model, uncertainty assessment, and dimming strategy updates are not performed independently, but rather through synergistic optimization.

[0231] Updates to the parameters of the causal model can affect the situational analysis results, thereby altering the input for uncertainty assessment. For example, updating the causal coefficient changes the predicted value of the target illuminance demand, requiring a recalculation of the corresponding prediction uncertainty. The system employs a cascading update mechanism: first, the causal model parameters are updated; then, the uncertainty is reassessed based on the new parameters; and finally, the dimming strategy is adjusted according to the new uncertainty assessment results.

[0232] During the update process, the system checks the consistency of each module after the update. The system uses a consistency verification mechanism to trigger a synchronized update of the corresponding module when such inconsistencies are detected.

[0233] Different components are updated at different frequencies. The confidence parameters of the data layer can be updated in real time; the causal coefficients of the model layer are updated in batches daily or weekly; and the weight configuration of the policy layer is optimized monthly through reinforcement learning. This layered management avoids system instability caused by overly frequent updates while ensuring timely correction of key parameters.

[0234] Through the aforementioned adaptive update mechanism, the system achieves continuous optimization across the entire chain, from data perception and model inference to decision execution.

Claims

1. A centralized control system for intelligent streetlights with multi-node adaptive dimming, characterized in that, include: The data interface layer is used to receive and aggregate multi-source operation data from multi-node smart streetlights, as well as external status data related to lighting needs. The lighting state construction module is used to construct an overall lighting state model in the spatial and temporal dimensions of a road or region based on the multi-source operational data and external state data. The overall lighting state model serves as an intermediate state representation of the system, used to uniformly characterize the lighting distribution state at different spatial locations, and as a shared state semantic input when multiple smart street light nodes make collaborative dimming decisions. The lighting situation analysis module is used to analyze and infer changes in lighting demand based on the overall lighting state model, and obtain situation analysis results that characterize the trend of changes in lighting demand. An uncertainty assessment module is used to assess the uncertainty of the external state data, the overall lighting state model, or the situation analysis results, and generate an uncertainty assessment result, wherein the uncertainty assessment result is used to characterize the credibility of the corresponding data or analysis results; The dimming decision and centralized control module is used to generate a collaborative dimming control strategy for multiple smart street light nodes based on the changes in lighting demand characterized by the lighting situation analysis results and under the credibility constraints defined by the uncertainty assessment results, and to distribute the collaborative dimming control strategy to the corresponding smart street light nodes for execution. The feedback and optimization module is used to obtain operational feedback information after dimming control is executed, and to update or adjust the overall lighting state model, uncertainty assessment results or dimming control strategy based on the operational feedback information to form closed-loop adaptive dimming control.

2. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 1, characterized in that: The multi-source operation data includes at least the node identifier, spatial location information, current brightness parameter or drive duty cycle parameter, energy consumption parameter and operation status information of the smart street light node; the external status data includes at least one of ambient light data, traffic flow data, personnel activity data or weather visibility data.

3. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 1, characterized in that: The overall lighting state model is a light field model used to describe the distribution of illuminance or brightness values ​​at different locations within a road or area.

4. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 3, characterized in that: The light field model is continuously updated over time to reflect the changing trend of the lighting state in the time dimension and to serve as the state input for dimming decisions.

5. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 1, characterized in that: The lighting situation analysis module analyzes the triggering conditions, scope of influence, or degree of influence of changes in lighting demand by constructing a causal relationship model among the factors influencing lighting demand; the causal relationship model is adaptively updated by the feedback and optimization module based on operational feedback information.

6. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 5, characterized in that: The causal relationship model is represented in the form of a structural causal model or a causal relationship graph, wherein the causal relationship edge configuration includes causal influence coefficients and time delay parameters.

7. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 1, characterized in that: The uncertainty assessment result is at least one of a confidence parameter, an uncertainty metric, or a probability distribution form.

8. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 7, characterized in that: The uncertainty assessment module quantifies uncertainty based on Bayesian inference methods, the statistical variance of prediction results, or the degree of discrepancy between prediction results from multiple models.

9. The intelligent street light centralized control system with multi-node adaptive dimming according to claim 1, characterized in that: When implementing collaborative dimming control, the dimming decision and centralized control module introduces at least one of lighting target constraints or energy consumption constraints. The dimming decision and centralized control module includes multiple functional software agents. Based on sharing the overall lighting state model, different software agents determine their respective dimming parameters through collaboration and game theory.

Images