Intelligent street lamp brightness adjustment method based on digital twinning

Abstract

The application discloses a kind of intelligence street lamp brightness adjustment methods based on digital twinning, to solve the problem that digital twin model is online calibrated and self-updating process is caused by lamp aging, pollution and individual difference, and long-term deviation of twin and entity, slow calibration convergence and active detection may affect lighting safety, the application is by constructing digital twin model and obtaining dimming instruction, measured illumination, measured electrical parameter and context information, generate optical residual and electrical residual;According to the context information, the hazard rate is adaptively generated, and the change point judgment is obtained by using Bayesian online change point detection, and the digital twin model and online sparse Gaussian process regression model are reset or segmented initialization when the change point triggers;Twin predicted illumination is used as the mean function of online sparse Gaussian process regression model to carry out residual online learning and compensation prediction, and under the illumination safety constraints such as minimum illumination threshold, maximum dimming change amplitude threshold and glare index threshold, the dimming instruction cycle update is selected and detected by using the context multi-arm Bandit based on Thompson Sampling, the technical effects of fast online calibration, long-term drift tracking and stable brightness adjustment under the illumination safety constraints are realized.

Description

Technical Field

[0001] This invention relates to the field of road lighting control technology, and in particular to a method for adjusting the brightness of smart streetlights based on digital twins. Background Technology

[0002] With the development of smart cities and the Internet of Things (IoT) in cities, LED streetlights are gradually acquiring the capabilities of remote dimming, status monitoring, and data reporting. Streetlight control has evolved from traditional timed switching to adaptive dimming control based on information such as time of day, traffic flow, pedestrian flow, ambient light, and weather. To improve the prediction and management capabilities of road lighting effects, some solutions introduce digital twin technology. By establishing a mapping relationship between factors such as streetlight installation parameters, light distribution characteristics, power supply parameters, and road reflections and illuminance output, simulation predictions of illuminance, electrical parameters, and glare under different dimming commands can be achieved. Simultaneously, to compensate for changes in the on-site environment and model errors, existing technologies also attempt to update model parameters using online learning or filtering methods, or identify status changes and trigger calibration through drift detection.

[0003] The existing technology still has the following shortcomings:

[0004] 1. Digital twin models usually rely on offline calibration or fixed parameters. When faced with aging lamps, pollution and individual differences, they are prone to long-term cumulative deviations. Online updates lack effective distinction between sudden and gradual changes, resulting in long-term deviations between the twin and the real entity and slow convergence speed.

[0005] 2. Drift or change point detection is mostly based on a single illuminance error or fixed prior, making it difficult to adaptively adjust the detection sensitivity by combining contextual information such as time period, weather, ambient light and temperature. This can easily lead to false alarms or missed alarms, resulting in frequent resets or missing key state changes.

[0006] 3. Active detection dimming, introduced to improve calibration efficiency, is prone to safety constraints such as minimum illuminance, brightness variation range, and glare in engineering applications. It lacks constraint screening and fallback strategies for feasible detection command sets, which poses a risk to lighting safety and user comfort.

[0007] Therefore, a method for adjusting street light brightness that can overcome the shortcomings of the prior art is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a method for adjusting the brightness of smart streetlights based on digital twins. Addressing the problems in existing technologies, such as the reliance on offline calibration of digital twin models leading to long-term deviations from the actual lighting due to aging, pollution, and individual differences in lamps; the difficulty in distinguishing between abrupt and gradual changes during online updates resulting in slow convergence and false / missed drift detection; and the lack of safety constraints in active dimming for calibration, potentially affecting road lighting safety, the following technical solution is proposed: A digital twin model is constructed and an online sparse Gaussian process regression model is initialized. Dimming commands, measured illuminance, measured electrical parameters, and contextual information are collected. Optical and electrical residuals are constructed based on twin predictions and measurements. A hazard rate is generated based on contextual information, and a Bayesian online change point detection method is used to determine the change point. When a change point is triggered, the parameters of the digital twin model and the sparse Gaussian process are reset or segmented initialized. The twin-predicted illuminance is used as the mean function of the sparse Gaussian process to learn the residuals online and perform compensation prediction. Under safety constraints such as minimum illuminance threshold, maximum dimming change threshold, and glare index threshold, a Thompson-based method is used... The sampling context uses a multi-arm bandit selection probe to iteratively update dimming commands. This invention achieves the technical advantages of rapid online calibration, long-term drift tracking, reduced false alarms, and improved brightness adjustment stability under lighting safety constraints.

[0009] This invention provides a method for adjusting the brightness of smart streetlights based on digital twins, comprising:

[0010] S1. Construct a digital twin model, initialize the online sparse Gaussian process regression model, and set lighting safety constraint parameters; S2. Obtain the current dimming command and collect measured illuminance, measured electrical parameters, and context information; S3. Input the current dimming command and context information into the digital twin model to obtain the twin predicted illuminance and twin predicted electrical parameters. Calculate the optical residual based on the measured illuminance and twin predicted illuminance, and calculate the electrical residual based on the measured electrical parameters and twin predicted electrical parameters; S4. Process the context information to generate a model representing the change point. S5. Input the optical residual, electrical residual, and risk rate into the Bayesian online change point detection to obtain the change point determination result; S6. If the change point determination result indicates that a change point has occurred, reset or segmentally initialize the model parameters of the digital twin model and the online sparse Gaussian process regression model; S7. Input the current dimming command and context information, using the measured illuminance as the observation, and the twin predicted illuminance as the mean function of the online sparse Gaussian process regression model to perform a calculation on the online sparse Gaussian process regression model. The process involves several steps: S1) updating the illuminance prediction posterior distribution, determining the residual posterior distribution of the optical residual using the illuminance prediction posterior distribution and the twin predicted illuminance, and outputting the residual prediction result; S8) generating multiple candidate dimming commands and inputting them into the reset or segmented initialized digital twin model with context information to obtain the corresponding twin predicted illuminance; inputting each candidate dimming command and context information into the reset or segmented initialized online sparse Gaussian process regression model, using the corresponding twin predicted illuminance as the mean function of the regression model to perform residual prediction, obtaining the illuminance prediction posterior distribution corresponding to each candidate dimming command, determining the residual prediction result corresponding to each candidate dimming command based on the illuminance prediction posterior distribution and the corresponding twin predicted illuminance, and performing residual compensation on the corresponding twin predicted illuminance based on the residual prediction result to obtain the compensated predicted illuminance; selecting a set of feasible detection dimming commands from multiple candidate dimming commands according to lighting safety constraint parameters, and selecting the detection dimming command to be issued from them; S9) issuing the detection dimming command to be issued to the street light for execution, and returning to step S2 at a preset sampling time.

[0011] Optionally, S1 includes:

[0012] When constructing the digital twin model, a parameterized mapping relationship is established between dimming commands and twin predicted illuminance, twin predicted electrical parameters, and twin predicted glare index, and the context information is used as one of the inputs to the parameterized mapping relationship. The parameterized mapping relationship is determined at least based on luminaire installation parameters, luminaire optical parameters, power supply parameters, and road surface parameters. The luminaire installation parameters include at least installation height and pole spacing; the luminaire optical parameters include at least light distribution characteristics and light output characteristics corresponding to the dimming command; the power supply parameters include at least input voltage; and the road surface parameters include at least reflectivity. Furthermore, model parameters for characterizing the effects of luminaire aging and pollution are set in the digital twin model.

[0013] When initializing the online sparse Gaussian process regression model, set the upper limit of the number of induced points, the initial position of the induced points, and the initial value of the kernel function hyperparameters;

[0014] When setting the lighting safety constraint parameters, at least the minimum illuminance threshold, the maximum dimming change threshold, and the glare index threshold should be set.

[0015] Terminology definition:

[0016] The digital twin model is a computational model used to characterize smart streetlights and their road lighting environment. It receives dimming commands and related inputs and outputs corresponding twin predicted illuminance, twin predicted electrical parameters and / or twin predicted glare index.

[0017] The model parameters used to characterize the effects of lamp aging and pollution are parameters used to describe the decrease in light output or changes in optical performance of lamps due to operating time or external pollution.

[0018] The online sparse Gaussian process regression model is a Gaussian process regression model that uses sparse approximation. During its operation, it is incrementally updated based on the online collected data to achieve prediction.

[0019] The induction point is a pseudo-input point or representative input point used to construct the sparse approximation of the online sparse Gaussian process regression model;

[0020] The upper limit of the number of induced points is the maximum number of induced points that the online sparse Gaussian process regression model is allowed to retain in the sparse approximation;

[0021] The initial position of the inducement point is the initial value or initial coordinate of the inducement point in the regression input space;

[0022] The kernel function hyperparameters are the parameters used in the kernel function of the online sparse Gaussian process regression model to determine the correlation structure and scale characteristics.

[0023] Optionally, S2 includes:

[0024] Obtain the current dimming command currently being executed by the smart street light, and synchronously sample the smart street light within a preset stable time period during which the current dimming command remains unchanged;

[0025] During synchronous sampling, the measured illuminance corresponding to the smart street light is collected by the illuminance acquisition device, the measured electrical parameters corresponding to the smart street light are collected by the electrical parameter acquisition module, and the context information corresponding to the road lighting area where the smart street light is located is collected by the sensing device and information interface.

[0026] The context information includes at least sampling time information, ambient light information, meteorological information, and lamp temperature information;

[0027] The measured illuminance, the measured electrical parameters, and the context information are timestamped to ensure that they correspond to the same sampling time.

[0028] Terminology definition:

[0029] The context information is a set of environmental and state information used to characterize the external and operational conditions related to road lighting at the sampling time.

[0030] The sampling time information is time information used to identify the time when the sampling occurred;

[0031] The ambient lighting information is lighting information used to characterize the intensity of natural light or the influence of surrounding light sources in the road lighting area;

[0032] The meteorological information refers to meteorological element information used to characterize the weather state at the time of sampling;

[0033] The timestamp alignment process involves appending or reading timestamps to the measured illuminance, the measured electrical parameters, and the context information, and then performing matching corrections to ensure that the three correspond to the same sampling time at the data level.

[0034] Optionally, S3 includes:

[0035] The current dimming command and context information are input into the digital twin model, and the digital twin model outputs the twin predicted illuminance and twin predicted electrical parameters corresponding to the current dimming command.

[0036] The optical residual is obtained by performing a difference operation between the measured illuminance and the twin predicted illuminance;

[0037] The measured electrical parameters and the twin predicted electrical parameters are compared using a differential calculation to obtain the electrical residuals;

[0038] The predicted electrical parameters and the measured electrical parameters include at least active power and input current. The optical residual is used to characterize the prediction deviation of the digital twin model for lighting output, and the electrical residual is used to characterize the prediction deviation of the digital twin model for electrical state.

[0039] Terminology definition:

[0040] The optical residual is the difference between the measured illuminance and the twin predicted illuminance, used to characterize the prediction error of the digital twin model for lighting output;

[0041] The electrical residual is the difference between the measured electrical parameter and the twin-predicted electrical parameter, used to characterize the prediction error of the digital twin model for the electrical state;

[0042] The difference operation is an operation that calculates the difference between corresponding variables at the same sampling time;

[0043] The active power is an electrical parameter that characterizes the effective electrical power actually consumed by the smart street light.

[0044] The input current is the electrical parameter of the current flowing into the input terminal or drive input terminal of the smart street light.

[0045] Optionally, S4 includes:

[0046] Feature extraction is performed on the context information to obtain a context feature vector. The context feature vector includes at least the time period feature corresponding to the sampling time information, the meteorological feature corresponding to the meteorological information, the ambient light feature corresponding to the ambient light information, and the temperature feature corresponding to the lamp temperature information.

[0047] The context feature vector is normalized, and the normalized context feature vector is mapped to a hazard rate based on a preset hazard rate mapping rule. The hazard rate mapping rule is used to make different time period features, meteorological features, ambient light features and temperature features correspond to different hazard rate values, so that the hazard rate changes adaptively with the context information.

[0048] Output the stated risk rate.

[0049] Terminology definition:

[0050] The risk rate is a parameter used to characterize the prior probability of a change point occurring at the current sampling time. It is used as a prior input for change point detection in Bayesian online change point detection and changes adaptively with context information.

[0051] The context feature vector is a vectorized representation obtained by extracting features from context information, and it includes at least one or more of time period features, meteorological features, ambient light features, and temperature features.

[0052] The feature extraction is a process of converting raw data in context information into feature quantities that can be used to calculate the risk rate.

[0053] The hazard rate mapping rule is a rule or functional relationship used to map the normalized context feature vector to a hazard rate, so that different context conditions correspond to different hazard rate values.

[0054] Optionally, S5 includes:

[0055] Optical and electrical residuals are constructed into residual observation vectors, and residual sequences are formed according to sampling times;

[0056] At each sampling time, the corresponding residual observation vector and hazard rate are input into the Bayesian online change point detection. The prior probability of the change point occurring is updated based on the hazard rate, and the posterior probability distribution of the run length is updated based on the residual sequence. The change point probability corresponding to the sampling time is calculated from the posterior probability distribution of the run length.

[0057] The change point probability is compared with a preset change point determination threshold. When the change point probability is not less than the change point determination threshold, a change point determination result indicating that a change point has occurred is generated. When the change point probability is less than the change point determination threshold, a change point determination result indicating that a change point has not occurred is generated.

[0058] Terminology definition:

[0059] The change point is the moment or time period boundary where the lighting output characteristics and / or electrical state characteristics of the smart street light undergo a statistical characteristic change in the time series;

[0060] The residual observation vector is a multidimensional observation formed by combining at least optical residuals and electrical residuals at the same sampling time.

[0061] The residual sequence is time series data composed of residual observation vectors arranged in the order of sampling time;

[0062] The Bayesian online change point detection is a detection method that updates the residual sequence online based on Bayesian inference and outputs the change point probability.

[0063] The prior probability of the change point occurring is the prior possibility of the change point occurring at that sampling moment before the introduction of residual observations at the current sampling moment.

[0064] The running length is the number of consecutive sampling steps without change points from the time of the most recent change point to the current sampling time.

[0065] The posterior probability distribution of the running length is the probability distribution of taking different values ​​of the running length after the residual sequence is observed;

[0066] The change point probability is a probability value calculated from the posterior probability distribution of the running length, used to characterize the possibility of a change point occurring at the current sampling time;

[0067] The change point determination threshold is a threshold parameter used to convert the change point probability into a change point determination result;

[0068] The change point determination result is a determination output that characterizes whether a change point has occurred at the current sampling time, and it includes at least two states: indicating that a change point has occurred and indicating that a change point has not occurred.

[0069] Optionally, S6 includes:

[0070] A segmentation flag is generated based on the change point determination result, wherein the segmentation flag includes a first state indicating that a change point has occurred and a second state indicating that no change point has occurred;

[0071] When the segmentation flag is in the first state, the sampling time of the change point is obtained, and the data segment corresponding to the digital twin model is divided into the data segment before the change point and the data segment after the change point based on the sampling time;

[0072] When the segmentation flag is in the first state, the model parameters of the digital twin model are reset or segmented initialized. The model parameters include at least the model parameters used to characterize the effects of lamp aging and pollution. The reset or segmented initialized digital twin model is then used to predict the data segment after the change point.

[0073] When the segmentation flag is in the first state, the sparse approximation parameters of the online sparse Gaussian process regression model are reset or segmented initialized. The sparse approximation parameters include at least the induced point position and variational parameters. The online sparse Gaussian process regression model after the reset or segmented initialization is then used to update the data segment after the point change online.

[0074] When the segmentation flag is in the second state, the model parameters of the digital twin model remain unchanged, and the sparse approximation parameters of the online sparse Gaussian process regression model remain unchanged.

[0075] Terminology definition:

[0076] The segmentation flag is an identifier used to characterize whether the current sampling time triggers the variable point segmentation process. It is used to control whether to perform a reset or segmentation initialization on the digital twin model and the online sparse Gaussian process regression model.

[0077] The first state is the state value in the segmentation flag used to indicate the occurrence of a change point;

[0078] The second state is the state value in the segmentation flag used to indicate that no change point has occurred;

[0079] The sampling time at which the change point occurs is the sampling time corresponding to the change point determination result indicating when the change point occurs, and is used as the time boundary for switching between data segmentation and model update strategies.

[0080] The data segment is a continuous set of data in a time series according to the sampling time;

[0081] The data segment before the change point is the historical data segment prior to the sampling time when the change point occurs;

[0082] The data segment after the change point refers to the sampling time at which the change point occurs and / or the subsequent data segment thereafter.

[0083] The model parameters are a set of parameters used to determine the mapping relationship between the output and input of the digital twin model;

[0084] The reset is a process that restores the model parameters to preset initial or default values ​​and prevents subsequent updates from inheriting the historical state before the change point;

[0085] The segmented initialization is to establish a new parameter initialization state for the data segment after the change point at the boundary of the change point, so that subsequent online updates are processed independently with the new data segment as the starting point.

[0086] The sparse approximation parameters are a set of parameters used to construct the sparse representation of the online sparse Gaussian process regression model;

[0087] The location of the induced point is the position value of the induced point in the regression input space, which is used to characterize the distribution of representative input points in the sparse approximation.

[0088] The variational parameters are used in sparse variational inference to characterize the approximate posterior distribution.

[0089] Optionally, the S7 includes:

[0090] The current dimming command and context information are used to construct a regression input vector. The measured illuminance is used as the observation, and the twin predicted illuminance is used as the mean function input to the reset or piecewise initialized online sparse Gaussian process regression model. Based on the regression input vector, the observation, and the mean function, sparse variational inference is used to update the variational parameters of the online sparse Gaussian process regression model online to obtain the illuminance prediction posterior distribution. The posterior distribution of the optical residual is determined by the illuminance prediction posterior distribution and the twin predicted illuminance, and the residual prediction result is output based on the residual posterior distribution. The residual prediction result includes at least the residual prediction mean and the residual prediction variance. An update weight coefficient is determined based on the electrical residual. When the electrical residual meets a preset abnormal condition, the update weight coefficient is set to a value less than the normal update weight coefficient, or the update weight coefficient is set to zero to freeze the update. During the online update, the update weight coefficient is used to weight the update contribution corresponding to the observation to obtain the updated online sparse Gaussian process regression model.

[0091] Terminology definition:

[0092] The regression input vector is a vectorized data composed of the current dimming command and the context information, and is used as the input to the online sparse Gaussian process regression model;

[0093] The observations are measured data used to drive the online update of the online sparse Gaussian process regression model, wherein the observations in step S7 are the measured illuminance;

[0094] The mean function is the prior mean term in the Gaussian process regression model, used to provide the baseline trend for prediction, wherein in step S7 the mean function is taken from the twin predicted illuminance;

[0095] The online update is a process of incrementally updating the parameters of the online sparse Gaussian process regression model at each sampling time using the newly obtained regression input vector and the observations;

[0096] The sparse variational inference is an inference method used to approximately obtain the posterior distribution by optimizing the variational objective function to update the variational parameters within the sparse approximation framework of the induced point.

[0097] The posterior distribution of the illuminance prediction is the probability distribution of the illuminance prediction value at the corresponding input point by the online sparse Gaussian process regression model after the regression input vector, the observations and the mean function are given and online updates are completed.

[0098] The residual posterior distribution is a probability distribution of the optical residual at the corresponding input, determined by the illuminance prediction posterior distribution and the twin predicted illuminance.

[0099] The residual prediction result is a residual statistic output based on the residual posterior distribution, including at least the residual prediction mean and the residual prediction variance;

[0100] The mean of the residual prediction is the expected value of the posterior distribution of the residual;

[0101] The residual prediction variance is the variance value of the residual posterior distribution, used to characterize the residual prediction uncertainty;

[0102] The update weight coefficient is a weight parameter used to adjust the contribution of the observation to the online update of the online sparse Gaussian process regression model;

[0103] The preset abnormal conditions are the conditional rules used to determine electrical abnormalities and adjust the updated weight coefficients accordingly.

[0104] The freeze update is achieved by setting the update weight coefficient to zero or an equivalent method so that the observations at the current sampling time do not participate in the model parameter update process;

[0105] The update contribution is the effective impact of the observation on the adjustment of model parameters during the online update.

[0106] The weighting is a process of scaling the update contribution proportionally using the update weight coefficient to suppress or enhance the impact of the update.

[0107] Optionally, S8 includes:

[0108] Multiple candidate dimming commands are generated based on the current dimming command, wherein the difference between each candidate dimming command and the current dimming command satisfies a preset dimming amplitude upper limit. Each candidate dimming command and context information are input into a reset or segmented initialized digital twin model, outputting the twin predicted illuminance and twin predicted glare index corresponding to each candidate dimming command. Each candidate dimming command and context information are input into a reset or segmented initialized online sparse Gaussian process regression model, using the corresponding twin predicted illuminance as the mean function of the online sparse Gaussian process regression model for residual prediction. The posterior distribution of illuminance prediction corresponding to each candidate dimming command is obtained by measurement; the residual prediction result corresponding to each candidate dimming command is determined by the posterior distribution of illuminance prediction and the corresponding twin predicted illuminance; the mean of the residual prediction in the residual prediction result is superimposed with the twin predicted illuminance corresponding to each candidate dimming command to obtain the compensated predicted illuminance corresponding to each candidate dimming command; based on the lighting safety constraint parameters, it is determined whether each candidate dimming command simultaneously satisfies the minimum illuminance threshold, the maximum dimming change threshold, and the glare index threshold, wherein the minimum illuminance threshold is determined by comparing the compensated predicted illuminance. The predicted illuminance is determined by comparing the twin predicted glare index with the glare index threshold. The maximum dimming change threshold is determined by calculating the change between the candidate dimming command and the current dimming command and comparing it with the maximum dimming change threshold. Candidate dimming commands that meet the lighting safety constraints are selected as a set of feasible detection dimming commands. In the set of feasible detection dimming commands, context information is used as context input, and a context-based multi-arm B-type algorithm based on ThompsonSampling is adopted. The ANDIT strategy selects the detection dimming command to be issued, wherein: the expected calibration benefit corresponding to each feasible detection dimming command is defined as its corresponding residual prediction variance; the actual calibration benefit is defined as the difference before and after the update of the residual prediction variance of the online sparse Gaussian process regression model at the corresponding input after executing the corresponding feasible detection dimming command and obtaining a new round of measured illuminance; the expected calibration benefit of each feasible detection dimming command is updated using Bayesian method based on the historical actual calibration benefit to obtain the posterior distribution of the expected calibration benefit, and Thompson Sampling is performed based on the posterior distribution to determine the detection dimming command to be issued; when the set of feasible detection dimming commands is empty, the current dimming command is determined as the detection dimming command to be issued.

[0109] Terminology definition:

[0110] The candidate dimming command is an alternative dimming control command generated based on the current dimming command for evaluation and selection;

[0111] The preset dimming amplitude upper limit is an upper limit parameter used to limit the dimming difference between the candidate dimming command and the current dimming command;

[0112] The dimming difference is the difference between the candidate dimming command and the current dimming command in terms of dimming level or dimming ratio.

[0113] The change amount is a value used to characterize the magnitude of the dimming change of the candidate dimming command relative to the current dimming command;

[0114] The compensated predicted illuminance is the compensated predicted value of the illuminance corresponding to the candidate dimming command, obtained by superimposing the residual prediction result on the twin predicted illuminance.

[0115] The lighting safety constraint parameters are a set of parameters used to determine the safety feasibility of the detection dimming command.

[0116] The minimum illuminance threshold is a constraint parameter used to limit the illuminance of road lighting areas from being lower than this threshold.

[0117] The maximum dimming change threshold is a constraint parameter used to limit the amount of dimming change in a single dimming operation from exceeding this threshold.

[0118] The glare index threshold is a constraint parameter used to limit the glare index from exceeding the threshold.

[0119] The feasible detection dimming command set is a set of candidate dimming commands that meet the safety constraints obtained by filtering multiple candidate dimming commands through the lighting safety constraint parameters.

[0120] The dimming command to be sent is a dimming command determined from the set of feasible dimming commands and prepared to be sent to the smart street light for execution.

[0121] The contextual multi-arm Bandit strategy is a decision-making strategy that selects an action from multiple candidate actions and updates it based on feedback under given contextual information, wherein each feasible detection dimming command corresponds to a Bandit arm.

[0122] Thompson Sampling is a decision-making method that randomly samples the posterior distribution of the returns of each candidate arm and selects the arm based on the sampling results.

[0123] The expected calibration gain is an indicator used to characterize the information gain expected at the calibration level from executing a feasible detection dimming command, and is defined as the residual prediction variance corresponding to the command.

[0124] The actual calibration benefit is the change in the residual prediction variance of the model at the corresponding input after executing a feasible detection dimming command and obtaining a new round of measured illuminance;

[0125] The Bayesian update is an update method that updates the probabilistic model parameters of the expected calibration benefit based on the historical actual calibration benefit to obtain the posterior distribution.

[0126] The posterior distribution is the probability distribution of the expected calibration gain obtained after incorporating historical actual calibration gain information.

[0127] Optionally, S9 includes:

[0128] The pending detection dimming command is sent to the controller of the smart street light, so that the controller adjusts the output brightness of the smart street light according to the pending detection dimming command;

[0129] After the detection dimming command to be issued is issued, a preset stabilization timer is started, and the acquisition operation of step S2 is triggered at the preset sampling time after the preset stabilization time ends, so as to obtain the measured illuminance, measured electrical parameters and context information corresponding to the detection dimming command to be issued.

[0130] The dimming command to be issued is used as the current dimming command in step S2 for the next iteration, thereby realizing online calibration and self-updating of the digital twin model and continuous adjustment of the brightness of the smart street light.

[0131] Optionally, an uncertainty adjustment factor is determined based on the residual prediction variance, and the uncertainty adjustment factor is used to adaptively adjust the detection intensity and change-point sensitivity, wherein: when the residual prediction variance is not less than a first variance threshold, the detection intensity and / or change-point sensitivity are increased, wherein increasing the detection intensity includes reducing the preset stabilization time in S2 and / or increasing the number of candidate dimming commands in S8, and increasing the change-point sensitivity includes increasing the hazard rate in S4 and / or decreasing the change-point determination threshold in S5; when the residual prediction variance is less than a second variance threshold, the detection intensity and / or change-point sensitivity are decreased, wherein decreasing the detection intensity includes increasing the preset stabilization time in S2 and / or decreasing the number of candidate dimming commands in S8, and decreasing the change-point sensitivity includes decreasing the hazard rate in S4 and / or increasing the change-point determination threshold in S5; wherein the first variance threshold is greater than the second variance threshold.

[0132] The beneficial effects of this invention are:

[0133] 1. By using twin-predicted illuminance as the mean function of an online sparse Gaussian process regression model, online learning is performed only on the optical residuals between the twin and the entity. The update weights are suppressed or frozen in combination with electrical residuals, thereby achieving rapid compensation for systematic biases caused by lamp aging, pollution, and individual differences. This reduces the sample size required for calibration and improves online convergence speed and long-term tracking capability.

[0134] 2. By simultaneously constructing optical and electrical residuals and adaptively generating hazard rate inputs based on contextual information, Bayesian online change point detection is achieved, enabling timely identification and differentiation of state changes such as power supply anomalies, drive changes, and sudden changes in optical performance, thereby reducing false alarms and missed alarms at change points. When a change point is triggered, the digital twin model and the online sparse Gaussian process regression model are reset or segmented initialized to avoid interference from historical data on the new state, thereby suppressing long-term deviations between the twin and the entity.

[0135] 3. By constructing a feasible set of detection dimming commands under lighting safety constraints such as minimum illuminance threshold, maximum dimming variation threshold, and glare index threshold, and by using Thompson Sampling-based contextual multi-arm bandit selection of detection dimming commands within the feasible set, we can achieve proactive acquisition of calibration information gain while ensuring lighting safety and comfort. Furthermore, when the feasible set is empty, we provide a fallback strategy to improve system robustness and feasibility. Attached Figure Description

[0136] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0137] Figure 1 This is a flowchart of a smart street light brightness adjustment method based on digital twin proposed in this invention;

[0138] Figure 2 This is a schematic flowchart illustrating the selection of active detection commands under security constraints in step S8 of the present invention. Detailed Implementation

[0139] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0140] refer to Figure 1 A method for adjusting the brightness of smart streetlights based on digital twins, comprising:

[0141] S1. Construct a digital twin model, initialize the online sparse Gaussian process regression model, and set lighting safety constraint parameters; S2. Obtain the current dimming command and collect measured illuminance, measured electrical parameters, and context information; S3. Input the current dimming command and context information into the digital twin model to obtain the twin predicted illuminance and twin predicted electrical parameters. Calculate the optical residual based on the measured illuminance and twin predicted illuminance, and calculate the electrical residual based on the measured electrical parameters and twin predicted electrical parameters; S4. Process the context information to generate a model representing the change point. S5. Input the optical residual, electrical residual, and risk rate into the Bayesian online change point detection to obtain the change point determination result; S6. If the change point determination result indicates that a change point has occurred, reset or segmentally initialize the model parameters of the digital twin model and the online sparse Gaussian process regression model; S7. Input the current dimming command and context information, using the measured illuminance as the observation, and the twin predicted illuminance as the mean function of the online sparse Gaussian process regression model to perform a calculation on the online sparse Gaussian process regression model. The process involves several steps: S1) updating the illuminance prediction posterior distribution, determining the residual posterior distribution of the optical residual using the illuminance prediction posterior distribution and the twin predicted illuminance, and outputting the residual prediction result; S8) generating multiple candidate dimming commands and inputting them into the reset or segmented initialized digital twin model with context information to obtain the corresponding twin predicted illuminance; inputting each candidate dimming command and context information into the reset or segmented initialized online sparse Gaussian process regression model, using the corresponding twin predicted illuminance as the mean function of the regression model to perform residual prediction, obtaining the illuminance prediction posterior distribution corresponding to each candidate dimming command, determining the residual prediction result corresponding to each candidate dimming command based on the illuminance prediction posterior distribution and the corresponding twin predicted illuminance, and performing residual compensation on the corresponding twin predicted illuminance based on the residual prediction result to obtain the compensated predicted illuminance; selecting a set of feasible detection dimming commands from multiple candidate dimming commands according to lighting safety constraint parameters, and selecting the detection dimming command to be issued from them; S9) issuing the detection dimming command to be issued to the street light for execution, and returning to step S2 at a preset sampling time.

[0142] In this specific embodiment, S1 includes:

[0143] A digital twin model is constructed using a modeling module deployed on a street light central controller or edge computing node, and the digital twin model uses dimming commands. With context information Input and output twin predicted illuminance Twin predictive electrical parameters and twin predictive glare index The context information The digital twin model includes at least sampling time information, ambient light information, meteorological information, and luminaire temperature information, organized in vector form as one of the inputs for parameterized mapping relationships. The static parameters of the digital twin model include at least luminaire installation parameters and environmental parameters, with the luminaire installation parameters including at least installation height. Distance from light pole The optical parameters of the luminaire include at least the light distribution characteristics. Light output characteristic curve corresponding to dimming command The power supply parameters include at least the input voltage. The road surface parameters include at least reflectivity. And in the digital twin model, model parameters are set to characterize the effects of lamp aging and pollution. ,in Used to characterize the combined effect of light output attenuation and light transmission loss, and its possible value range in implementation is: Its initial value Choose to set it to 1 to indicate the initial no-decay state and allow it to be overridden by a reset or segmented initialization in subsequent steps to adapt to the new state after the change point;

[0144] In this specific embodiment, the "parametric mapping relationship between dimming commands and twin predicted illuminance" is constructed primarily based on physical mechanism calculations and secondarily by configurable correction factors. Specifically, a set of illuminance evaluation points is preset within the road lighting area and based on the installation height... Spacing between light poles , light distribution characteristics and road surface reflectivity Calculations are obtained in the dimming command Baseline illuminance ,in In practice, this can be achieved by performing geometric propagation calculations on the incident light rays at each evaluation point and combining the light distribution intensity with the cosine of the incident angle. Furthermore, it can be calculated based on the spacing between light poles. Determine the contribution range of adjacent luminaires to obtain the evaluation point illuminance for single or multiple lamps, and further incorporate contextual information. Mapped to environment correction function With meteorological attenuation function ,in This method is used to characterize the impact of luminaire temperature information on luminous flux maintenance and driving efficiency, and can be implemented using piecewise linear or lookup table methods. This is used to characterize the light propagation attenuation caused by meteorological information such as fog, rain, and snow, and can be obtained by looking up tables from visibility or rainfall intensity, ultimately yielding twin predicted illuminance. The calculation method is as follows:

[0145] ;

[0146] in Indicates dimming command With context information Twin-predicted illuminance, These represent the model parameters used to characterize the effects of lamp aging and pollution. Indicates context information A defined temperature correction function, Indicates context information A defined meteorological attenuation function, Indicates based on dimming command Installation height Spacing between light poles , light distribution characteristics road surface reflectivity The calculated baseline illuminance;

[0147] The "parameterized mapping relationship between dimming commands and twin predictive electrical parameters" in its implementation requires at least the output active power and input current, along with the input voltage of the power supply parameters. Consistent, specifically based on dimming commands. The rated current model is obtained from the drive load curve. Then based on context information Obtain the electrical efficiency correction function To characterize the impact of temperature and power supply fluctuations on drive efficiency, and thus obtain the twin-predicted input current. Twin-based prediction of active power ,in Efficiency curves segmented by lamp temperature or a lookup table method can be used, and configuration is allowed depending on the product model. The twin predictive electrical parameters... At least including and It can also be expanded to include electrical parameters such as power factor, and the selected calculation method is:

[0148] ;

[0149] in Indicates twin-predicted input current. Indicates dimming command A defined rated current model, Indicates context information A defined electrical efficiency correction function, Indicates the twin-predicted active power. Indicates the input voltage;

[0150] The "parametric mapping relationship between dimming commands and twin-predicted glare indices" is constructed using a glare evaluation method consistent with road lighting engineering. Specifically, it involves pre-setting the observer's position and line of sight, and basing the method on light distribution characteristics. Installation height Spacing between light poles And the illuminance predicted by twins The calculated road surface background brightness estimate is used to calculate the glare contribution and output the twin-predicted glare index. The glare index can be selected using a uniform glare value (UGR) or an equivalent glare index, and the selection is fixed through a configuration file. The road surface reflectivity... The ambient lighting information is used for background brightness estimation to ensure that glare assessment is consistent with road materials, while the ambient lighting information serves as contextual information. A portion of it is used to overlay and correct the background brightness to reflect the effect of ambient light or moonlight on glare perception;

[0151] After completing the construction of the digital twin model, the online sparse Gaussian process regression model is initialized. Specifically, the upper limit of the number of induced points is preset in the implementation. To limit the computational and storage overhead of online updates, the initial set of induced point positions is initialized within the regression input space. ,in and Consistent with subsequent steps, the input vector is obtained by sampling a regression vector consisting of dimming commands and context information, and the dimming command is selected as the input vector. Allowed range and context information Generate by performing Latin hypercube sampling or uniform grid sampling within the normalized feature range. To cover common working conditions, the initial values ​​of the kernel function hyperparameters are set to ensure that the model can converge. The kernel function is selected to be a squared exponential kernel or a Matérn kernel, and its hyperparameters include at least the initial values ​​of signal variance, length scale, and observation noise variance. The initial values ​​can be obtained from the variance statistics of historical calibration data or configured according to product experience values.

[0152] After completing the model initialization, set the lighting safety constraint parameters, specifically including at least the minimum illuminance threshold. Maximum dimming variation threshold and glare index threshold ,in This is used to ensure that the predicted illuminance for subsequent compensation is not lower than the safety lower limit required for road lighting levels, and can be configured according to road category. It is used to constrain the amount of change of a single dimming command relative to the current dimming command to avoid sudden perception by the driver and can be configured according to the control cycle. This is used to constrain the twin-predicted glare index to not exceed the comfort limit and to be consistent with the selected glare index type.

[0153] In this specific embodiment, S2 includes:

[0154] The data acquisition module first obtains the current dimming command being executed by the smart street light through the readback interface of the smart street light controller. ,in For the first The dimming command corresponding to the iteration cycle can be either a dimming level or a dimming duty cycle. The readback interface is selected as DALI readback. One or more of sampling readback or PWM duty cycle readback are used to ensure that the "actual effective instruction" is obtained rather than just the issued instruction;

[0155] After confirmation Then the stability control logic is started, making Preset stable duration The interior remains unchanged, where This is used to cover the drive response and LED thermal steady-state establishment process and can be configured in the configuration file according to the luminaire model. If a change in the dimming command is detected, the data for that round is discarded and the system re-enters a stable timing loop to ensure that the current round of synchronous sampling corresponds to the same current dimming command.

[0156] In the Synchronous sampling is performed during this period, and a unified system clock is used to generate the sampling time sequence. ,in Indicates the first Round The sampling time of the next sample and For sampling sequence number, For the first The number of samplings within a stable duration is obtained by the sampling period configuration. The system clock selection is provided by the central controller or edge node and the clock is synchronized between the sensing device and the electrical parameter acquisition module via NTP or PTP to reduce cross-device time deviation.

[0157] At each sampling time The illuminance acquisition device collects the original measured illuminance value corresponding to the smart street light and completes the calibration conversion to obtain the illuminance sample. At the same time, the electrical parameter acquisition module collects the measured electrical parameter sample corresponding to the smart street light. The measured electrical parameters include at least active power and input current, and are selected to be provided by the metering chip or the built-in metering unit of the driver to ensure accuracy and synchronization.

[0158] The system collects contextual information samples corresponding to the road lighting area where the smart street light is located through sensing devices and information interfaces. The contextual information includes at least sampling time information, ambient light information, meteorological information, and lamp temperature information. The sampling time information is directly given by the system clock. The ambient light information is selected to be collected by the ambient light sensor. The meteorological information is selected to be obtained by the local meteorological sensor or the meteorological service interface. The lamp temperature information is selected to be collected by the lamp body temperature sensor or the driver temperature sensor.

[0159] To meet the data consistency requirements of "synchronous sampling," the measured illuminance samples, the measured electrical parameter samples, and the context information samples are time-stamp aligned, i.e., using... The timestamps of each data source are matched using the reference sampling time, and the "nearest neighbor matching + tolerance threshold" method is selected to address the time tolerance issue. Inner selection and The closest sample is used as the alignment sample when a data source is in If no samples are available, linear interpolation is used to generate aligned samples. If it is still impossible to generate them, the sampling time is marked as invalid and removed from subsequent statistics to avoid erroneous data driving model updates.

[0160] After completing timestamp alignment, to reduce the impact of instantaneous noise and occasional interference on subsequent residual calculations, the aligned samples within the stable time period are aggregated to obtain the measured illuminance. Measured electrical parameters and context information Determine using the following formula:

[0161] ;

[0162] in, Indicates the first The measured illuminance obtained from rounds of iteration, Indicates the first The measured electrical parameters obtained from the rounds of iteration include at least active power and input current. Indicates the first The context information vector obtained from the round of iterations, Indicates the first The effective number of samplings within a stable duration of each iteration. Indicates the sampling sequence number. Indicates at the sampling time Illuminance samples aligned with timestamps Indicates at the sampling time The timestamp-aligned electrical parameter sample vector This represents the timestamp-aligned context information vector corresponding to the last valid sampling time within the stable duration, and is used to characterize the context conditions during this round of sampling. This indicates the iteration round index.

[0163] In this specific embodiment, S3 includes:

[0164] The prediction and residual calculation module will use the current dimming command and context information As a unified input for the same iteration, the digital twin model is fed into the forward calculation. The digital twin model uses the current model parameters (including model parameters used to characterize the effects of lamp aging and pollution) that have been constructed and loaded in step S1. The parameterized mapping relationship is used to output twin predicted illuminance that corresponds one-to-one with the input condition. and twin-predicted electrical parameters To ensure the comparability of measured and predicted illuminance, the twin predicted illuminance... The calculation points are selected within the digital twin model according to the same spatial location and installation orientation as the illuminance acquisition device. Specifically, the installation coordinates, pointing angle, and ground clearance of the illuminance acquisition device are configured as output sampling points for the digital twin model and fixed in the model configuration file, ensuring that the output illuminance matches the measured illuminance. For the same sampling location, the twin predicted electrical parameters It includes at least active power and input current, and is combined with the input voltage in step S1. Once the power supply parameters are consistent, a difference operation is then performed on the measured values ​​and the twin predictions to construct residual observations for subsequent steps, specifically calculating the optical residuals. and electrical residual The optical residual is used to characterize the prediction deviation of the digital twin model for lighting output, and the electrical residual is used to characterize the prediction deviation of the digital twin model for electrical state. In the implementation, the electrical residual is selected as a two-dimensional vector to simultaneously characterize the two channels of active power and input current, thereby providing a more stable anomaly criterion in subsequent change point detection and updated weight control.

[0165] The difference operation uses signed differences to preserve the direction of the deviation, and its calculation method is as follows:

[0166] ;

[0167] in, Indicates the first Optical residuals of round iterations, Indicates the first Measured illuminance in each iteration, Indicates the current dimming command With context information The following is the twin-predicted illuminance output by the digital twin model. Indicates the first The electrical residual vector of the round of iterations, Indicates the first The measured electrical parameter vector of the round of iterations, which includes at least active power. With input current Indicates the current dimming command With context information The following is a vector of predicted electrical parameters output by the digital twin model, which includes at least the predicted active power. With twin prediction of input current This indicates the iteration round index.

[0168] In this specific embodiment, S4 includes:

[0169] The hazard rate generation module uses context information The input and output are the hazard rates used to characterize the prior probability of a change point occurring. ,in It should include at least the sampling time information, ambient light information, meteorological information, and lamp temperature information, and For iteration round index;

[0170] right Perform feature extraction to form contextual feature vectors ,in At least by time period characteristics Meteorological feature vectors Ambient lighting characteristics and temperature characteristics The time period features are obtained by splicing. The meteorological feature vector is constructed by selecting indicators such as "number of minutes within the day" and "whether it is a dusk or dawn window" determined by sampling time information to enhance sensitivity to transition periods between light switching on and off. The method of using unique thermal coding of meteorological types and superimposing continuous quantities such as visibility or rainfall intensity is selected to distinguish operating conditions such as fog, rain, and snow, which are more prone to abrupt changes in light propagation and electrical states. The ambient lighting characteristics are described below. The ambient light sensor output can be selected and its short-term rate of change can be superimposed to reflect disturbances from external light sources such as vehicle lights or billboards, as described in the temperature characteristic. Select to collect lamp temperature information and superimpose the temperature rise rate to reflect sudden changes in thermal state;

[0171] Subsequently Normalization is performed to eliminate dimensional differences and ensure the reusability of the hazard rate mapping rule. The normalization method chosen is min-max normalization along the feature dimension to obtain the normalized context feature vector. The calculation method is as follows:

[0172] ;

[0173] in express The Normalization characteristics, express The Dimensional original features, Indicates the first The minimum value of the dimensional feature determined in the engineering configuration or historical statistics. Indicates the first The maximum value of a dimensional feature determined in engineering configuration or historical statistics. Indexed by feature dimensions, For iteration round index;

[0174] In obtaining Then, based on the preset hazard rate mapping rules, Mapped to risk rate The hazard rate mapping rule is used to assign different hazard rate values ​​to different time period characteristics, meteorological characteristics, ambient light characteristics, and temperature characteristics, and to make the hazard rate adaptively change with context information. A bounded logical mapping with configurable weights is selected to ensure... It also facilitates rapid deployment on different roads and lighting fixtures, and can be specifically determined using the following formula:

[0175] ;

[0176] in Indicates the first Risk rate of round iteration This represents the lower bound of the hazard rate to avoid the change point never occurring due to a prior value of zero, and its value is chosen to be... to This represents the upper limit of the risk rate to avoid excessively high prior values ​​leading to frequent false alarms, and its value is chosen to be between 0.05 and 0.3. This represents the Sigmoid function. This represents the independent variable of the Sigmoid function. This represents the bias coefficient. Indicates and A weight vector with consistent dimensions. This indicates the transpose operation. This represents the normalized context feature vector. Represents the natural constant;

[0177] in and It can be obtained offline by fitting historical operating data with "change point determination result" as the label or configured by engineering experience. The option is to assign higher weights to dusk / dawn windows, severe weather, rapid changes in ambient light, and high temperature or sudden temperature rise of lamps to improve the sensitivity of change point detection under these conditions.

[0178] Output the aforementioned risk rate .

[0179] In this specific embodiment, S5 includes:

[0180] The change point detection module receives optical residuals. With electrical residuals and risk rate The optical and electrical residuals are then constructed into residual observation vectors to form a residual sequence recursively based on the sampling time.

[0181] Specifically, electrical residual Decomposed into active power channel residuals Residual of input current path And construct the residual observation vector for:

[0182] ;

[0183] in Indicates the first The residual observation vector of the round of iterations, Indicates the iteration round index. Indicates the first Optical residuals of round iterations, Indicates the first Electrical residuals of active power channels in each iteration Indicates the first Electrical residuals of the input current path in round iterations. Indicates the transpose operation;

[0184] Then according to The residual sequence is incrementally generated and input into the Bayesian online change point detection to maintain the posterior probability distribution of the runtime length online, where the runtime length is a random variable. In the The value of the round iteration and Used to characterize "from the most recent change point to the [number]th change point". The number of consecutive, unchanging point sampling steps up to the end of the iteration is truncated to ensure real-time operation. ,in Candidate values ​​for runtime. The maximum reserved value for the running length can be configured from 200 to 2000 according to the device's computing power;

[0185] In each iteration, Bayesian online change point detection uses a hazard rate. As a priori input to "the current round is a turning point", that is, in the case of the first round... Wheel direction When iterating through rounds, use probability Transferred to and with probability Transferred to And combined with residual observation vector Prediction likelihood under various runtime assumptions Update runtime posterior distribution ,in Indicates the first The runtime during round iteration is taken as: The posterior probability, This indicates that the current running length is "Under this assumption, the observed..." The predicted probability;

[0186] The recursive update is performed according to the following formula, and after obtaining the unnormalized result, [the following is done]... according to Normalize the expression to make its sum equal to 1, and choose to compute in the logarithmic field and use log-sum-exp to avoid numerical underflow:

[0187] ;

[0188] in Indicates the first The posterior probability distribution of the runtime of each iteration is in The probability value at that location. Indicates the first The posterior probability distribution of the runtime of each iteration is in The probability value at that location. This indicates the run-length index used for summation. Indicates the first The danger rate generated by the round iteration based on contextual information and This represents the residual observation vector under the corresponding run length assumption. The prediction seems likely. This represents the summation operation. This indicates a proportional relationship, with the right side representing the unnormalized probability. Indicates the maximum retention value for the runtime;

[0189] in In the implementation, the observation model is constructed using the principle that "the residual observation vectors within each segment are independent and identically distributed, and the conditions of each channel are independent." Each channel is treated as a one-dimensional Gaussian distribution, and a conjugate prior is used to maintain a sufficient statistic online for each channel, thereby obtaining a one-dimensional Student-t prediction distribution for each channel. The prediction probabilities of each channel are then multiplied together to obtain the final value. and in Prior sufficient statistics are used to initialize the changepoint branches to reflect the "start of a new segment";

[0190] In obtaining The probability of the change point was then determined as follows: The probability of the change point is compared with a preset change point determination threshold. When comparing, The output indicates the change point determination result when a change point occurs. The output indicates the result of the change point determination that has not occurred.

[0191] In this specific embodiment, S6 includes:

[0192] The model management module receives the change point determination results and generates segmentation markers accordingly. The segmentation flag is used to control whether to perform a reset or segmented initialization on the digital twin model and the online sparse Gaussian process regression model, and is directly assigned a value from the change point determination result as follows:

[0193] ;

[0194] in Indicates the first The segmentation flag corresponding to the round iteration. This represents the iteration round index, where 1 indicates the first state (i.e., a change point has occurred) and 0 indicates the second state (i.e., no change point has occurred).

[0195] when When the change point occurs, obtain the sampling time and select the current round sampling time directly in the implementation. As the sampling time when the change point occurs, The sampling time formed in step S2 and used for timestamp alignment, subsequently... To define the boundaries, the data segments corresponding to the digital twin model are divided into pre-change data segments and post-change data segments, and a segmentation index is established at the data level. Specifically, this involves... Mark the data segment before the change point and The data segment is marked as the data segment after the change point, where For the first The current dimming command in the round iteration, For the first Measured illuminance in each iteration, For the first Measured electrical parameters of each iteration For the first Context information of round iteration, Index for historical rounds;

[0196] After data segmentation, the model parameters of the digital twin model are reset or initialized in segments. The model parameters of the digital twin model are denoted as follows: And it should include at least model parameters used to characterize the effects of lamp aging and pollution. The reset strategy is to Restore to factory or initial configuration parameters and at least Set to initial value and The value is set to 1 to ensure that the prediction after the change point does not inherit the attenuation state before the change point. The segmented initialization strategy keeps structural parameters such as lamp installation parameters and road surface parameters unchanged and only initializes parameters strongly correlated with aging and pollution. and the calibrable correction factor is reinitialized, and the The initial values ​​for the segmentation were estimated by the ratio of measured illuminance to twin baseline illuminance within the first stable observation window after the change point, and then truncated and restricted to... To satisfy the physical meaning;

[0197] Simultaneously, the sparse approximation parameters of the online sparse Gaussian process regression model are reset or piecewise initialized, and the sparse approximation parameters include at least the induced point locations. With variational parameters ,in Represents the set of induced point locations. Indicates the first The location of each induced point is within the regression input space consisting of dimming commands and context information. Indicates the current number of induced points and satisfies This indicates the upper limit of the number of induction points set in step S1. The variational parameters used for sparse variational inference are represented and chosen to be composed of the variational mean vector and variational covariance matrix of the induced variables;

[0198] Under the reset strategy, Z is regenerated according to the initial position generation rules of the induced point in step S1 and... To restore the prior state and clear the residual memory learned before the change point, under the segmented initialization strategy, The most recent regression input vectors from the data segment after the change point are selected and generated through clustering or uniform sampling to ensure that the induced points cover the input distribution in the new state, while maintaining the kernel function hyperparameters at their latest values ​​before the change point to reduce reconvergence time. To avoid numerical instability, initialization is performed using an approximate posterior-oriented prior-centered approach, employing the following variational distribution form and initial value setting: ;

[0199] in This indicates that the online sparse Gaussian process regression model affects the inducing variables. The variational posterior distribution, Indicates a Gaussian distribution. Indicates the location at the induction point Let m be the vector of function values ​​at point M, and S be the variational mean vector. Represents the zero vector. This indicates the position of the kernel function at the induced point. The kernel matrix obtained from the above calculation;

[0200] when At the same time, maintain the model parameters of the digital twin model. Invariant and maintain the sparse approximation parameters of the online sparse Gaussian process regression model and This remains unchanged to achieve continuous online updates under conditions of no change points.

[0201] In this specific embodiment, S7 includes:

[0202] The online learning module will display the current dimming command. With context information The vectors are concatenated and normalized according to the same rules as in step S1 to form the regression input vector. and the measured illuminance The online sparse Gaussian process regression model is reset or piecewise initialized as the observation input to perform online updates;

[0203] The online sparse Gaussian process regression model is implemented using the "illuminance function" as the regression object and employing a sparse variational approximation, where the regression input is... The set of induced point locations is And the number of inducement points does not exceed the upper limit of the number of inducement points. The variational parameter is Furthermore, in the implementation, the variational mean vector of the induced variables is chosen. With variational covariance matrix The kernel function is constructed using the kernel function and its hyperparameters set in step S1, and is consistent with the segmentation strategy.

[0204] To implement the residual learning mechanism of "using twin-predicted illuminance as the mean function", the prior of the online sparse Gaussian process regression model is set as follows:

[0205] ;

[0206] in This indicates that the regression model is in response to the input. The corresponding illuminance function value at that location, This represents the regression input vector consisting of dimming commands and context information. Represents a Gaussian process. The mean function of a Gaussian process. The kernel function representing a Gaussian process. Indicates and Another regression input vector of the same dimension, This indicates that the dimming command is generated by a digital twin model. With context information The twin-predicted illuminance output below, This indicates a dimming command. Represents a context information vector;

[0207] Under this prior setting, sparse variational inference is used to update the variational parameters during online updates. Perform incremental updates, specifically based on the current sample. The variational lower bound is updated once or multiple times using stochastic gradient optimization based on the Gaussian observation noise assumption to obtain the updated value. and The observation noise variance is taken as the initial value of the observation noise variance in the hyperparameter of the kernel function and can be configured in engineering according to the accuracy of the acquisition device.

[0208] To suppress the contamination of residual learning by electrical anomalies, based on electrical residuals... Determine the updated weight coefficients ,in And under normal circumstances, it is set to 1. When the electrical residual meets the preset abnormal conditions, it will be... The update is frozen by setting the value to a value less than the normal update weight coefficient or setting it to 0. The preset abnormal conditions include, in implementation, "the absolute value of the electrical residual of the active power channel exceeds a threshold" and / or "the absolute value of the electrical residual of the input current channel exceeds a threshold." The threshold can be configured according to a fixed ratio of the drive's rated power to its rated current, and two levels of thresholds can be set to correspond to "weighted update" and "frozen update" respectively. During online optimization... The likelihood contribution of the current sample is weighted to suppress or mask the updated contribution.

[0209] After the online update is completed, the output of the online sparse Gaussian process regression model is... The posterior distribution of illuminance at a given location is predicted and its mean is used in the implementation. With variance Characterization, in which This represents the posterior mean of the posterior distribution of illuminance prediction. This represents the posterior variance of the posterior distribution of the illuminance prediction.

[0210] The posterior distribution of the optical residuals is determined by the predicted illuminance posterior distribution and the twin predicted illuminance, and the residual prediction result is output. Specifically, the residual is defined as "the deviation of illuminance from the twin prediction", and the mean of the residual prediction is obtained. With residual prediction variance as follows:

[0211] ;

[0212] in Indicates the first The mean of the residual prediction in each iteration. Indicates the first The posterior mean of illumination prediction in each iteration. This indicates the digital twin model in the current dimming command. With context information Twin-predicted illuminance, Indicates the first The residual prediction variance of each iteration Indicates the first Posterior variance of illumination prediction in round iterations Indicates the first The current dimming command in the round iteration, Indicates the first The context information vector of each iteration, This indicates the iteration round index.

[0213] In this specific embodiment, S8 includes:

[0214] The active detection decision module uses the current dimming command and context information Generate multiple candidate dimming command sets based on the baseline. ,in For the iteration round index, The generation selection adopts "with The dimming command is pre-configured within a symmetrical discrete sampling pattern centered on the target and a boundary truncation strategy is applied. With candidate interval ,according to Generate and exceed Eliminate items, and force all candidates to satisfy the condition. The difference does not exceed the maximum dimming change threshold. To meet the requirements for smoothness in single dimming;

[0215] For each candidate dimming command The two-stage forecasting process of "twin forecasting + residual forecasting + residual compensation" is executed separately, with the first stage being... and The twin predicted illuminance is obtained by inputting the reset or segmented initialization digital twin model. and twin predictive glare index Then and The constructed regression input vector is input to the online sparse Gaussian process regression model after reset or piecewise initialization, and then... The residual prediction is obtained by using it as a mean function. With residual prediction variance The residual prediction mean is used to compensate for the twin predicted illuminance to obtain the compensated predicted illuminance. Based on lighting safety constraint parameters, a set of feasible detection dimming commands was obtained. It is calculated and determined according to the following formula:

[0216] ;

[0217] in Indicates candidate dimming command In context information The corresponding compensated predicted illuminance, This represents the twin-predicted illuminance output by the digital twin model. This represents the mean of the residual predictions output by the online sparse Gaussian process regression model. This represents the set of feasible detection dimming commands that satisfy lighting safety constraints. This represents the set of candidate dimming commands. Indicates the minimum illuminance threshold. This indicates the threshold for the maximum dimming variation. Indicates the glare index threshold. This represents the twin-predicted glare index output by the digital twin model. Indicates the current dimming command. This represents the absolute value operation. Represents a set;

[0218] In obtaining Then, with context information As contextual input, a contextual multi-armed bandit strategy based on Thompson Sampling is employed. Select the detection dimming command to be sent. Each feasible detection dimming command will be included. Treat it as a Bandit arm and define its expected calibration gain as the residual prediction variance corresponding to that arm. The actual calibration benefit is defined as the difference between the predicted variance of the residual at the input and the updated value after executing the arm and completing the next round of online updates. To achieve repeatable engineering feasibility, a Gaussian posterior distribution of the expected calibration gain is maintained for each arm. With fixed observation noise variance Perform a Bayesian update, where Configure system parameters and retrieve historical data. The variance estimate, Thompson Sampling sampling and selection, and the posterior update are implemented as follows:

[0219] ;

[0220] in Indicates from the first Round The expected calibration return posterior distribution of the bar arm is used to sample the return samples. Indicates a Gaussian distribution. Indicates the first Round The posterior mean of the expected calibration benefit of the bar arm. Indicates the first Round The posterior variance of the expected calibration benefit of the bar arm. Indicates the first The selected wheel yields the detection and dimming commands to be sent. This represents the independent variable that maximizes the objective. This represents the variance of the observed noise, indicating the actual calibration benefit. Indicates the first Round of execution The actual calibration benefit obtained after the arm is selected and calculated as the difference between the updated value and the original value of the residual prediction variance at the corresponding input point of that arm. This represents the operation of reciprocal of the square. and They represent the first The updated posterior mean and updated posterior variance for each round;

[0221] when If empty, the current dimming command will be executed. Confirmed as awaiting issuance of detection and dimming command This is to establish a rollback strategy and prevent the system from being unable to output control commands due to safety constraints.

[0222] In this specific embodiment, S9 includes:

[0223] The execution and closed-loop scheduling module will receive the detection dimming command. The data is sent to the controller of the smart street light via a communication link. Indicates the first The detection and dimming commands to be issued are selected through rounds of iteration. This indicates the iteration round index, and the communication link is selected as DALI, LoRa, NB-IoT, or... One of the methods employs a reliable interaction process of "send, read back, and verify," meaning the controller receives and executes... The returned execution confirmation frame and the actual effective dimming value are then sent back. If the returned actual effective dimming value is different from the actual effective dimming value... If there is a discrepancy, it will be recorded as a failure to send the command and a retry will be triggered, or the previous dimming command will be retained to avoid abnormal jumps.

[0224] Record the time of this distribution after confirming successful distribution. It also starts a preset stable duration timer, and at the same time uses a locking strategy to ensure that the controller remains stable within the stable window. The condition remains unchanged to satisfy synchronous sampling, where Indicates the first The instruction issuance time for each iteration is preset to a stable duration of [time value]. Furthermore, consistent with the preset stabilization time used in step S2 to ensure the stability of optical and electrical states, the controller within the stabilization window can use ramp limiting or gradual change strategies to achieve a smooth transition of brightness, but it must ensure that a steady state is reached at the end of the stabilization window.

[0225] To ensure a one-to-one correspondence between sampling data and detection dimming commands, the data buffers of the illuminance acquisition device, electrical parameter acquisition module, and sensing device are simultaneously activated during the stabilization window and timestamped according to a unified system clock. At the end of the stabilization window, alignment and packaging are triggered, and this moment is set as the preset sampling moment for the next round of acquisition. It satisfies:

[0226] ;

[0227] in Indicates the first The first sampling moment of the round iteration is used to trigger the acquisition operation in step S2. Indicates the first The timing of issuing instructions for round iteration. Indicates the preset stable duration;

[0228] exist When step S2 is triggered, The current dimming command in step S2 is written into the current iteration context and used for subsequent timestamp alignment and residual calculation. This ensures that the measured illuminance, measured electrical parameters, and context information output in step S2 are strictly bound to the detected dimming command. Furthermore, to maintain the consistency of iteration variables, the current dimming command for the next iteration is recorded as:

[0229] ;

[0230] in Indicates the first The current dimming command in the round iteration, Indicates the first The iteration is awaiting the issuance of detection and dimming commands. Indicates the iteration round index;

[0231] After completing the above distribution, stabilization, and triggering, the system enters the next cycle and returns to step S2 to realize the online calibration and self-update of the digital twin model and the continuous adjustment of the brightness of the smart street light.

[0232] In this specific embodiment, it also includes:

[0233] The system is additionally configured with an uncertainty adaptive adjustment module and performs parameter self-tuning once after completing step S7 in each round. The residual prediction variance is used as a quantitative index of model uncertainty and the detection intensity and change point sensitivity are adjusted accordingly.

[0234] Specifically, the adjustment module reads the residual prediction variance output in step S7. And based on the first variance threshold With the second variance threshold Calculate the uncertainty adjustment factor ,in Used to characterize high uncertainty boundaries that require "increased detection intensity and / or increased sensitivity to change points". The two thresholds, used to characterize the low uncertainty boundary that requires "reduced detection intensity and / or reduced change point sensitivity", can be set offline by the quantiles of the historical residual prediction variance or configured by engineering experience according to road grade and sensing accuracy.

[0235] The uncertainty adjustment factor Choose to calculate based on the truncated linear mapping to ensure... ,when hour Setting it to 1 indicates maximizing detection intensity and sensitivity to changes in point; when hour Setting it to 0 indicates minimizing detection intensity and sensitivity at variable points; the calculation method is as follows:

[0236] ;

[0237] in Indicates the first Uncertainty adjustment factor for round iteration, Indicates the iteration round index. Indicates the first The residual prediction variance output in step S7 of the round iteration. This represents the first variance threshold. This represents the second variance threshold. This indicates the minimum value operation. This indicates the operation of finding the maximum value.

[0238] In obtaining Then, it is used to adaptively adjust the key parameters of steps S2, S4, S5 and S8 in the next iteration to achieve a trade-off between closed-loop stability and convergence speed. The adjustment of detection intensity includes at least adjusting the preset stabilization time of step S2 and the number of candidate dimming commands in step S8. The adjustment of change point sensitivity includes at least adjusting the danger rate output in step S4 and the change point judgment threshold used in step S5. In order to avoid exceeding the scope of engineering implementation, all adjusted parameters are set with upper and lower bounds and truncated according to the upper and lower bounds.

[0239] Specifically, the preset stabilization time used in the next round step S2 is set to And limited to Inside, among which The minimum stabilization time should be no less than the sum of the drive response time and the sampling alignment buffer time. To maximize the stabilization time and reduce detection frequency and communication overhead under low uncertainty, the number of candidate dimming commands generated in the next step S8 is set to [value missing]. And limited to Inside, among which This is used to ensure that at least one basic fallback candidate exists. Used to limit computational load and must be consistent with the upper limit of the number of induced points. And matching edge computing power;

[0240] Simultaneously, in the next round of execution step S4, the basic hazard rate will be obtained first according to the hazard rate mapping rule. Then, an uncertainty-driven scaling is performed to obtain the adjusted hazard rate. The scaling factor Set as Monotonically increasing and satisfying To achieve a reduction in risk rate under low uncertainty and an increase in risk rate under high uncertainty, and to... Cut off to the upper and lower limits of the risk rate In order to maintain its validity as a prior probability;

[0241] Simultaneously, in the next execution of step S5, the change point determination threshold will be updated to... And limited to Inside, among which A smaller threshold corresponds to higher change point sensitivity. A larger threshold corresponds to a lower change point sensitivity;

[0242] The above adjustment is achieved using the following formula:

[0243] ;

[0244] in Indicates the first The preset stabilization time used in iteration step S2 is... This indicates the lower bound of the preset stable duration. This indicates the upper bound of the preset stable duration. Indicates the first The number of candidate dimming commands generated in iteration step S8. This indicates the lower bound of the number of candidate dimming commands. This indicates the upper bound of the number of candidate dimming commands. This indicates the floor function. Indicates by the first Uncertainty adjustment factor A defined risk rate scaling factor, This indicates the lower bound of the hazard rate scaling factor. This indicates the upper bound of the hazard rate scaling factor. Indicates the first Round iteration according to step S4 based on context information The base risk rate obtained from the mapping, Indicates the first The adjusted hazard rate is input in round iteration step S5. This indicates the lower limit of the risk rate. This indicates the upper limit of the risk rate. Indicates the first The threshold for determining the change point is used in iteration step S5. This indicates the lower limit of the threshold for determining a change point. This indicates the upper limit of the threshold for determining the change point;

[0245] In the above way, when Not less than Automatically decrease and increase To increase detection intensity and simultaneously increase and reduce To improve sensitivity to change points, when Less than Automatically increases and reduce To reduce detection intensity and simultaneously reduce and increase To reduce the sensitivity to changes in light intensity, and the above adjustments do not change the minimum illuminance threshold. Maximum dimming variation threshold Glare index threshold This ensures that lighting safety constraints remain effective at all times.

[0246] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0247] This invention addresses the problem of "long-term deviation between digital twin and physical entity" caused by aging of lamps, pollution, and changes in power supply and drive status during long-term operation of smart streetlights. It constructs a closed-loop calibration mechanism consisting of digital twin prediction, residual modeling, change point detection, and active safety detection. Specifically, the digital twin model provides baseline predictions under different dimming commands and contextual conditions. Measured illuminance and electrical parameters, along with the twin output, form optical and electrical residual sequences. Bayesian online change point detection performs online inference on the residual sequences under the prior constraint of contextual hazard rate to identify state abrupt changes and trigger model reset or segmented initialization, avoiding continuous contamination of new states by historical data. The online sparse Gaussian process regression model updates the residuals online and outputs the posterior distribution of the residuals, thereby compensating for twin predictions and continuously compressing the deviation between the twin and the actual entity. Simultaneously, by screening feasible dimming command sets under lighting safety constraints such as minimum illuminance, maximum dimming variation amplitude, and glare, and employing a contextual multi-arm bandit strategy based on Thompson Sampling to select detection commands, it achieves proactive acquisition of high-information samples at a lower cost, improving calibration speed and reducing long-term drift risk while ensuring lighting safety.

[0248] This invention addresses the aforementioned technical problems with scenario-oriented structural improvements: First, it uses twin-predicted illuminance as the mean function of the sparse Gaussian process regression model, enabling the regression model to learn only the residuals rather than the full illuminance mapping, thereby improving online convergence speed and reducing sample requirements. Second, it uses optical and electrical residuals as dual-channel observations and introduces a context-adaptive hazard rate, allowing change point detection to adjust sensitivity based on time period, weather, ambient light, and temperature, reducing false alarms and missed alarms. Third, the change point-triggered reset or segmented initialization mechanism directly uses the detection results to switch the update strategy between digital twin parameters and sparse approximation parameters, structurally suppressing the accumulation of long-term deviations. Fourth, the feasible set construction and fallback strategy under safety constraints make active detection engineering-ready, ensuring lighting safety and comfort while continuously improving model calibration quality, thus more effectively achieving stable adjustment of smart street light brightness and long-term consistency of the digital twin.

Claims

1. A digital-twin-based intelligent street light brightness adjustment method, characterized in that, include: S1. Construct a digital twin model, initialize the online sparse Gaussian process regression model, and set lighting safety constraint parameters; S2. Obtain the current dimming command and collect the measured illuminance, measured electrical parameters, and context information; S3. Input the current dimming command and context information into the digital twin model to obtain the twin predicted illuminance and twin predicted electrical parameters. Calculate the optical residual based on the measured illuminance and twin predicted illuminance, and calculate the electrical residual based on the measured electrical parameters and twin predicted electrical parameters. S4. Extract features from the context information to obtain a context feature vector. The context feature vector includes at least the time period feature corresponding to the sampling time information, the meteorological feature corresponding to the meteorological information, the ambient light feature corresponding to the ambient light information, and the temperature feature corresponding to the lamp temperature information. The context feature vector is normalized, and the normalized context feature vector is mapped to a hazard rate based on a preset hazard rate mapping rule. The hazard rate mapping rule is used to make different time period features, meteorological features, ambient light features and temperature features correspond to different hazard rate values, so that the hazard rate changes adaptively with the context information. Output the stated risk rate; S5. Input the optical residual, electrical residual, and hazard rate into the Bayesian online change point detection to obtain the change point determination result; S6. If the change point determination result indicates that a change point has occurred, reset or perform segmented initialization of the model parameters of the digital twin model and the online sparse Gaussian process regression model. S7. Input the current dimming command and context information, use the measured illuminance as the observation, use the twin predicted illuminance as the mean function of the online sparse Gaussian process regression model to update the online sparse Gaussian process regression model online, obtain the illuminance prediction posterior distribution, determine the residual posterior distribution of the optical residual by the illuminance prediction posterior distribution and the twin predicted illuminance, and output the residual prediction result. S8. Generate multiple candidate dimming commands based on the current dimming command, wherein the difference between each candidate dimming command and the current dimming command satisfies a preset dimming amplitude upper limit. Each candidate dimming command and context information is input into the reset or segmented initialization digital twin model, and the twin predicted illuminance and twin predicted glare index corresponding to each candidate dimming command are output. Each candidate dimming command and context information is input into the online sparse Gaussian process regression model after reset or segmented initialization. The corresponding twin predicted illuminance is used as the mean function of the online sparse Gaussian process regression model to perform residual prediction, thereby obtaining the posterior distribution of illuminance prediction corresponding to each candidate dimming command. The residual prediction result corresponding to each candidate dimming command is determined by the posterior distribution of each illuminance prediction and the corresponding twin predicted illuminance; The mean of the residual prediction in the residual prediction result is superimposed with the twin predicted illuminance corresponding to each of the candidate dimming commands to obtain the compensated predicted illuminance corresponding to each of the candidate dimming commands. Based on the lighting safety constraint parameters, it is determined whether each candidate dimming command simultaneously meets the minimum illuminance threshold, the maximum dimming change threshold, and the glare index threshold, and the candidate dimming commands that meet the lighting safety constraint parameters are selected as a set of feasible detection dimming commands. In the set of feasible detection dimming commands, context information is used as context input, and a context-based multi-arm bandit strategy based on Thompson Sampling is adopted to select the detection dimming command to be issued, wherein: The expected calibration gain corresponding to each of the feasible detection dimming commands is defined as its corresponding residual prediction variance. The actual calibration benefit is defined as the difference before and after updating the residual prediction variance of the online sparse Gaussian process regression model at the corresponding input after executing the corresponding feasible detection dimming command and obtaining a new round of measured illuminance. Based on historical actual calibration benefits, the expected calibration benefits of each feasible detection dimming command are updated using Bayesian method to obtain the posterior distribution of the expected calibration benefits, and Thompson Sampling is performed based on the posterior distribution to determine the detection dimming command to be issued. When the set of feasible detection dimming commands is empty, the current dimming command is determined as the detection dimming command to be issued; S9. Send the dimming command to be sent to the streetlight for execution, and return to step S2 at the preset sampling time, including: An uncertainty adjustment factor is determined based on the residual prediction variance, and this uncertainty adjustment factor is used to adaptively adjust the detection intensity and change-point sensitivity, wherein: When the residual prediction variance is not less than the first variance threshold, the detection intensity is increased and / or the change point sensitivity is increased. The increase in detection intensity includes reducing the preset stabilization time and / or increasing the number of candidate dimming commands. The increase in change point sensitivity includes increasing the risk rate and / or reducing the change point determination threshold. When the residual prediction variance is less than the second variance threshold, the detection intensity is reduced and / or the change point sensitivity is reduced. The reduction of detection intensity includes increasing the preset stabilization time and / or reducing the number of candidate dimming commands. The reduction of change point sensitivity includes reducing the danger rate and / or increasing the change point determination threshold. Wherein, the first variance threshold is greater than the second variance threshold.

2. The digital-twin-based intelligent street light brightness adjustment method according to claim 1, wherein S2 include: Obtain the current dimming command currently being executed by the smart street light, and synchronously sample the smart street light within a preset stable time period during which the current dimming command remains unchanged; During synchronous sampling, the measured illuminance corresponding to the smart street light is collected by the illuminance acquisition device, the measured electrical parameters corresponding to the smart street light are collected by the electrical parameter acquisition module, and the context information corresponding to the road lighting area where the smart street light is located is collected by the sensing device and information interface. The context information includes at least sampling time information, ambient light information, meteorological information, and lamp temperature information; The measured illuminance, the measured electrical parameters, and the context information are timestamped to ensure that they correspond to the same sampling time.

3. The digital-twin-based intelligent street light brightness adjustment method according to claim 1, wherein S3 includes: The current dimming command and context information are input into the digital twin model, and the digital twin model outputs the twin predicted illuminance and twin predicted electrical parameters corresponding to the current dimming command. The optical residual is obtained by performing a difference operation between the measured illuminance and the twin predicted illuminance; The measured electrical parameters and the twin predicted electrical parameters are compared using a differential calculation to obtain the electrical residuals; The predicted electrical parameters and the measured electrical parameters include at least active power and input current. The optical residual is used to characterize the prediction deviation of the digital twin model for lighting output, and the electrical residual is used to characterize the prediction deviation of the digital twin model for electrical state.

4. The digital-twin-based intelligent street light brightness adjustment method according to claim 1, characterized in that S5 include: Optical and electrical residuals are constructed into residual observation vectors, and residual sequences are formed according to sampling times; At each sampling time, the corresponding residual observation vector and hazard rate are input into the Bayesian online change point detection. The prior probability of the change point occurring is updated based on the hazard rate, and the posterior probability distribution of the run length is updated based on the residual sequence. The change point probability corresponding to the sampling time is calculated from the posterior probability distribution of the run length. The change point probability is compared with a preset change point determination threshold. When the change point probability is not less than the preset change point determination threshold, a change point determination result indicating that a change point has occurred is generated. When the change point probability is less than the preset change point determination threshold, a change point determination result indicating that no change point has occurred is generated.

5. The digital-twin-based intelligent street light brightness adjustment method according to claim 1, characterized in that S6 include: A segmentation flag is generated based on the change point determination result, wherein the segmentation flag includes a first state indicating that a change point has occurred and a second state indicating that no change point has occurred; When the segmentation flag is in the first state, the sampling time of the change point is obtained, and the data segment corresponding to the digital twin model is divided into the data segment before the change point and the data segment after the change point based on the sampling time; When the segmentation flag is in the first state, the model parameters of the digital twin model are reset or segmented initialized. The model parameters include at least the model parameters used to characterize the effects of lamp aging and pollution. The reset or segmented initialized digital twin model is then used to predict the data segment after the change point. When the segmentation flag is in the first state, the sparse approximation parameters of the online sparse Gaussian process regression model are reset or segmented initialized. The sparse approximation parameters include at least the induced point position and variational parameters. The online sparse Gaussian process regression model after the reset or segmented initialization is then used to update the data segment after the point change online. When the segmentation flag is in the second state, the model parameters of the digital twin model remain unchanged, and the sparse approximation parameters of the online sparse Gaussian process regression model remain unchanged.

6. The digital-twin-based intelligent street light brightness adjustment method according to claim 1, wherein S7 includes: The current dimming command and context information are used to construct a regression input vector. The measured illuminance is used as the observation, and the twin predicted illuminance is used as the mean function input of the online sparse Gaussian process regression model after reset or piecewise initialization. Based on the regression input vector, the observations, and the mean function, sparse variational inference is used to update the variational parameters of the online sparse Gaussian process regression model online to obtain the posterior distribution of illumination prediction. The posterior distribution of the optical residual is determined by the predicted posterior distribution of illuminance and the twin predicted illuminance, and the residual prediction result is output based on the posterior distribution of the residual. The residual prediction result includes at least the residual prediction mean and the residual prediction variance. The update weight coefficient is determined based on the electrical residual, wherein when the electrical residual meets the preset abnormal conditions, the update weight coefficient is set to a value less than the normal update weight coefficient, or the update weight coefficient is set to zero to freeze the update. When performing the online update, the update contribution corresponding to the observation is weighted using the update weight coefficient to obtain the updated online sparse Gaussian process regression model.

7. The method for adjusting the brightness of a smart street light based on digital twins according to claim 1, characterized in that, S9 includes: The pending detection dimming command is sent to the controller of the smart street light, so that the controller adjusts the output brightness of the smart street light according to the pending detection dimming command; After the detection dimming command to be issued is issued, a preset stabilization timer is started, and the acquisition operation of step S2 is triggered at the preset sampling time after the preset stabilization time ends, so as to obtain the measured illuminance, measured electrical parameters and context information corresponding to the detection dimming command to be issued. The dimming command to be issued is used as the current dimming command in step S2 for the next iteration, thereby realizing online calibration and self-updating of the digital twin model and continuous adjustment of the brightness of the smart street light.

Images