Machine learning-based downlight energy-saving dimming intelligent decision method and system

Abstract

The application discloses a kind of energy-saving dimming intelligent decision-making method and system based on machine learning. By disturbing each down lamp, the influence coefficient matrix is constructed according to the response of illumination sampling point;Using the matrix and natural light component, the expected illumination of dimming amount combination is solved, the environmental characteristics are input, the combination that meets the illumination and the lowest energy consumption is taken as a label, and the dimming decision model is trained.When running, the model outputs dimming amount and controls down lamp;According to the deviation between actual illumination and expected illumination, the matrix is corrected, and the training sample is regenerated, and the new and old samples are mixed according to the matrix correction amplitude;When the number of deviation threshold accumulation reaches the preset number, continue training based on the existing weight, and use the updated model for subsequent dimming.The application drives training sample reconstruction and model parameter update through illumination contribution relationship drift, improves the adaptive updating ability of dimming decision model to down lamp illumination contribution relationship drift.

Description

Technical Field

[0001] This application relates to the field of supervised training and adaptive update technology of machine learning models, specifically to a machine learning-based intelligent decision-making method and system for energy-saving dimming of downlights. Background Technology

[0002] Intelligent decision-making for energy-saving dimming of downlights typically utilizes environmental information such as population distribution, time of day, and natural lighting to determine the dimming level of each downlight, reducing energy consumption while meeting the illuminance requirements of each illuminance sampling point. When using a machine learning model, environmental features can be used as model input, and the dimming levels of each downlight corresponding to different lighting scenarios can be used as supervision labels. The trained model then outputs dimming decisions during the runtime phase.

[0003] Existing data-driven dimming methods typically train dimming decision models using fixed historical samples. As downlights age, their installation status changes, or the spatial environment changes, the contribution of a unit dimming amount to the illuminance at each sampling point deviates from the correspondence established during training. The target dimming amount corresponding to similar environmental characteristics also changes accordingly. If the model parameters maintain the mapping of the original operating conditions, the output dimming amount may result in insufficient local illuminance, or the overall dimming output may be increased to compensate for the deviation. Relying on re-accumulating large amounts of historical data for periodic training delays model parameter updates, leading to insufficient adaptive updating capability of the dimming decision model to the drift in the downlight illuminance contribution relationship. Summary of the Invention

[0004] This application provides a machine learning-based intelligent decision-making method and system for energy-saving dimming of downlights, which can improve the adaptive update capability of the dimming decision model to the drift of the downlight illuminance contribution relationship.

[0005] In a first aspect, this application provides a machine learning-based intelligent decision-making method for energy-saving dimming of downlights, comprising the following steps: The dimming level of each downlight is changed while keeping the dimming level of the remaining downlights constant. An influence coefficient matrix is ​​constructed based on the changes in dimming level and the changes in illuminance at each illuminance sampling point. Multiple sets of environmental feature information are obtained, and the expected illuminance of the dimming amount combination is calculated using the influence coefficient matrix and the natural light illuminance component. The environmental feature information is used as the sample input, and the target dimming amount combination that meets the target illuminance and has the lowest energy consumption is used as the sample label. An initial training sample set is generated, and a machine learning model is trained based on the initial training sample set to obtain a dimming decision model. The current environmental characteristics information is input into the dimming decision model, which outputs the target dimming amount for each downlight and controls the dimming of each downlight. Based on the total illuminance of each illuminance sampling point before dimming, the current dimming amount of each downlight, and the actual illuminance after dimming, the natural light illuminance component, expected illuminance, and illuminance deviation are determined using the influence coefficient matrix. The influence coefficient matrix is ​​then corrected based on the changes in dimming amount and illuminance. The training samples are regenerated using the corrected influence coefficient matrix. The mixing ratio of the old and new training samples is determined according to the matrix correction magnitude to form an updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point to continue training using the updated sample set, thereby obtaining an updated dimming decision model for subsequent dimming operations.

[0006] Optionally, the dimming level of each downlight is changed while keeping the dimming level of the remaining downlights constant. An influence coefficient matrix is ​​constructed based on the changes in dimming level and the changes in illuminance at each illuminance sampling point. Specifically, this includes: Select a single downlight as the disturbance object according to a preset order, keep the dimming amount of the other downlights unchanged, and make the dimming amount of the disturbance object jump from the current value to another dimming amount within a short time window, and simultaneously collect the illuminance values ​​of each illuminance sampling point before and after the jump. Determine the illuminance change at each illuminance sampling point before and after the jump, divide the illuminance change at each illuminance sampling point by the dimming amplitude of the dimming, and use the illuminance influence coefficient corresponding to the disturbance object. Go through each downlight one by one, and arrange the illuminance influence coefficients according to the correspondence between the illuminance sampling points and the downlights to obtain the influence coefficient matrix.

[0007] Optionally, multiple sets of environmental feature information are obtained, and the expected illuminance of the dimming combination is calculated using the influence coefficient matrix and the natural light illuminance component. The initial training sample set is generated using the environmental feature information as sample input and the target dimming combination that satisfies the target illuminance while minimizing energy consumption as sample labels. Specifically, this includes: Multiple sets of personnel distribution characteristics, time period characteristics, and natural lighting characteristics were obtained respectively. Each set of characteristics was normalized and spliced ​​to obtain the corresponding environmental characteristic information and natural light illuminance components. For each set of environmental feature information, the expected illuminance under multiple dimming combinations is calculated using the influence coefficient matrix and the corresponding natural light illuminance component. From the dimming combination that satisfies the target illuminance corresponding to each illuminance sampling point, the target dimming combination with the lowest dimming energy consumption is determined. Each set of environmental feature information is associated with the corresponding target dimming amount combination to form a training sample. The sample input of the training sample is the environmental feature information, and the sample label is the target dimming amount combination. Multiple sets of training samples are collected to obtain the initial training sample set.

[0008] Optionally, training a machine learning model based on the initial training sample set to obtain a dimming decision model specifically includes: The initial training sample set is divided into a training set and a validation set; Using the training set, with sample inputs as model inputs and sample labels as supervised outputs, the model parameters of the machine learning model are iteratively updated through a loss function; Training stops when the loss function value corresponding to the validation set meets the preset convergence condition or reaches the preset number of iterations, thus obtaining the dimming decision model.

[0009] Optionally, inputting current environmental characteristic information into the dimming decision model, outputting the target dimming amount for each downlight, and controlling the dimming of each downlight specifically includes: Real-time data collection of current personnel distribution characteristics, time period characteristics, and natural lighting characteristics is used, and the current environmental feature information is obtained by adopting the same feature composition and normalization method as in the model training phase. The current environmental characteristic information is input into the dimming decision model, and the target dimming amount combination consisting of the target dimming amount of each downlight is output. A dimming control command is generated based on the target dimming amount combination, and the dimming control command is sent to each downlight to control each downlight to perform dimming.

[0010] Optionally, based on the total illuminance at each illuminance sampling point before dimming, the current dimming level of each downlight, and the actual illuminance after dimming, the influence coefficient matrix is ​​used to determine the natural light illuminance component, expected illuminance, and illuminance deviation. The influence coefficient matrix is ​​then corrected based on the changes in dimming level and illuminance. Specifically, this includes: Before dimming is performed, record the total illuminance at each illuminance sampling point and the current dimming level of each downlight, and collect the actual illuminance when the illuminance stabilizes after dimming is performed. Subtract the product of the influence coefficient matrix and the vector corresponding to the current dimming amount from the total illuminance to obtain the natural light illuminance component. Use the influence coefficient matrix, the target dimming amount, and the natural light illuminance component to calculate the expected illuminance and the illuminance deviation. When the illuminance deviation exceeds the preset deviation threshold, the influence coefficient matrix is ​​corrected by using the changes in dimming amount and illuminance before and after multiple dimming operations as observations, and by employing a recursive least squares algorithm or a batch least squares algorithm.

[0011] Optionally, training samples are regenerated using the corrected influence coefficient matrix. The mixing ratio of the old and new training samples is determined according to the matrix correction magnitude to form an updated sample set. When the cumulative number of times the illuminance deviation exceeds a preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as a starting point to continue training using the updated sample set, resulting in an updated dimming decision model for subsequent dimming operations. Specifically, this includes: The process of re-calculating the expected illuminance and determining the target dimming amount combination is re-executed using the corrected influence coefficient matrix to generate new training samples. The matrix correction magnitude is determined based on the difference before and after the matrix correction. The mixing ratio of the new training samples and the original training samples is determined according to the matrix correction magnitude. The larger the matrix correction magnitude, the higher the mixing ratio of the new training samples, thus obtaining the updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the training starting point, and the updated sample set is used to continue training to obtain the updated dimming decision model, which is then used for subsequent dimming operations.

[0012] Secondly, this application provides a machine learning-based intelligent decision-making system for energy-saving dimming of downlights, comprising: The response modeling module is used to change the dimming level of each downlight while keeping the dimming level of the remaining downlights constant, and to construct an influence coefficient matrix based on the changes in dimming level and the changes in illuminance at each illuminance sampling point. The initial training module is used to acquire multiple sets of environmental feature information, calculate the expected illuminance of the dimming amount combination using the influence coefficient matrix and the natural light illuminance component, use the environmental feature information as sample input, and the target dimming amount combination that meets the target illuminance and has the lowest energy consumption as sample label to generate an initial training sample set and train a machine learning model based on the initial training sample set to obtain a dimming decision model. The decision execution module is used to input the current environmental feature information into the dimming decision model, output the target dimming amount of each downlight, and control the dimming of each downlight; The drift observation module is used to determine the natural light illuminance component, expected illuminance, and illuminance deviation based on the total illuminance of each illuminance sampling point before dimming, the current dimming amount of each downlight, and the actual illuminance after dimming, using the influence coefficient matrix. The influence coefficient matrix is ​​then corrected based on the changes in dimming amount and illuminance. The model update module is used to regenerate training samples using the corrected influence coefficient matrix, determine the mixing ratio of the old and new training samples according to the matrix correction magnitude, and form an updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point, and the updated sample set is used to continue training to obtain the updated dimming decision model for subsequent dimming operation.

[0013] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned intelligent decision-making method for energy-saving dimming of downlights based on machine learning.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described intelligent decision-making method for energy-saving dimming of downlights based on machine learning.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This application establishes a physically meaningful influence coefficient matrix by perturbing each lamp individually, and uses the influence coefficient matrix and natural light illuminance components to calculate the expected illuminance under different lighting scenarios. The target dimming amount combination that satisfies the target illuminance and minimizes dimming energy consumption is used as a supervision label, thereby generating an initial training sample set and training the dimming decision model. During the operation phase, the target dimming amount output by the model is sent to the actual downlights. The actual illuminance after dimming constitutes independent feedback and is used to correct the influence coefficient matrix on which the generated samples are based. The matrix correction magnitude changes the composition of the old and new training samples. The accumulated illuminance deviation triggers continued training based on the existing weights, so that the physical response drift is sequentially incorporated into the training data composition, model parameter update, and subsequent actual lamp control. This reduces the need to re-accumulate historical operating data after changes, reduces the impact of the deviation of the downlight illuminance contribution relationship on the dimming decision, and improves the adaptive update capability of the dimming decision model to the drift of the downlight illuminance contribution relationship. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a machine learning-based intelligent decision-making method for energy-saving dimming of downlights, according to some embodiments of this application. Figure 2 This is a schematic flowchart illustrating the process of determining a target dimming amount combination according to some embodiments of this application; Figure 3 This is a schematic diagram of the control dimming process according to some embodiments of this application; Figure 4This is a schematic diagram of the structure of a machine learning-based intelligent decision-making system for energy-saving dimming of downlights, as shown in some embodiments of this application. Figure 5 This is an internal structural diagram of a computer device that implements a machine learning-based intelligent decision-making method for energy-saving dimming of downlights, according to some embodiments of this application. Detailed Implementation

[0017] To better understand the technical solutions in this embodiment, the technical solutions in this embodiment will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0018] refer to Figure 1 The figure is a flowchart illustrating a machine learning-based intelligent decision-making method for energy-saving dimming of downlights according to some embodiments of this application. This machine learning-based intelligent decision-making method for energy-saving dimming of downlights mainly includes the following steps: In step 101, the dimming amount of each downlight is changed while keeping the dimming amount of the remaining downlights constant. An influence coefficient matrix is ​​constructed based on the changes in dimming amount and the changes in illuminance at each illuminance sampling point.

[0019] In some embodiments, the dimming level of each downlight is changed while the dimming level of the remaining downlights remains constant. The construction of an influence coefficient matrix based on the changes in dimming level and the changes in illuminance at each illuminance sampling point specifically includes: Select a single downlight as the disturbance object according to a preset order, keep the dimming amount of the other downlights unchanged, and make the dimming amount of the disturbance object jump from the current value to another dimming amount within a short time window, and simultaneously collect the illuminance values ​​of each illuminance sampling point before and after the jump. Determine the illuminance change at each illuminance sampling point before and after the jump, divide the illuminance change at each illuminance sampling point by the dimming amplitude of the dimming, and use the illuminance influence coefficient corresponding to the disturbance object. Go through each downlight one by one, and arrange the illuminance influence coefficients according to the correspondence between the illuminance sampling points and the downlights to obtain the influence coefficient matrix.

[0020] In practice, the disturbance identification process can be initiated at night or during periods when no one is indoors to avoid interference from dimming fluctuations. The system selects one downlight in the space as the disturbance target according to a preset sequence, keeping the dimming levels of the other downlights constant. The dimming level of the selected downlight changes from its current value to a preset amplitude within a short window. In this embodiment, the preset amplitude can be 20% of the full dimming range, and the window duration can be one to three seconds. Since this duration is much shorter than the time scale of natural light changes, it can be approximated that the natural light remains constant during the disturbance. Before and after the amplitude change, the illuminance values ​​of each illuminance sampling point are collected synchronously, and the difference in illuminance before and after the amplitude change at the same sampling point is taken as the illuminance change caused by the disturbance. Preferably, the disturbance is repeated multiple times for the same downlight with the same amplitude, and the average value of the illuminance change obtained from each disturbance is taken. To suppress the influence of random noise from illuminance sensing on the identification results; further, if the change in natural light during the disturbance is not negligible, a reference sampling point unaffected by the illumination of each downlight can be set in the space, and the illuminance drift of the reference sampling point before and after the disturbance can be collected synchronously, and the drift can be subtracted from the illuminance change of each illuminance sampling point, thereby reducing the influence of natural light drift on the identification results; for each downlight, the illuminance change of each illuminance sampling point caused by the disturbance is compared with the dimming change of the downlight, and the ratio is the illuminance influence coefficient of the downlight on each illuminance sampling point; after traversing all downlights in the space one by one in the above manner, the influence coefficients corresponding to each downlight are arranged according to the correspondence between the illuminance sampling point and the downlight, and an influence coefficient matrix characterizing the contribution of each downlight to the illuminance of each illuminance sampling point is constructed.

[0021] It should be noted that the disturbance dimming described in this application refers to a calibration operation that, while keeping the operating state of the other downlights unchanged, causes a short-term, small jump in the dimming amount of a single downlight, and collects the illuminance response at each illuminance sampling point caused by this jump. Its purpose is to separate the contribution of each downlight to the spatial illuminance distribution from the total illuminance of the multiple lights. The illuminance sampling points refer to illuminance collection units pre-arranged at different locations within the space where the downlights are located. Their number and location can be determined according to the functional zoning of the space and the requirements for illuminance uniformity; for example, they can be arranged at office workstations and corridors. Representative locations such as windowsills; the influence coefficient matrix refers to a numerical matrix composed of each illuminance sampling point as a row and each downlight as a column, where any matrix element represents the amount of illuminance change caused by the unit dimming amount change of the corresponding downlight at the corresponding illuminance sampling point. Its physical meaning is to instantiate the physical law of linear superposition of illuminance of each light source under incoherent light source illumination into a solvable numerical model, so that the total illuminance of any illuminance sampling point can be expressed as the sum of the illuminance contributed by each downlight and the natural light illuminance, thereby providing a physical basis for the forward calculation of the expected illuminance and the correction of illuminance deviation.

[0022] In step 102, multiple sets of environmental feature information are obtained, and the expected illuminance of the dimming amount combination is calculated using the influence coefficient matrix and the natural light illuminance component. The environmental feature information is used as the sample input, and the target dimming amount combination that meets the target illuminance and has the lowest energy consumption is used as the sample label to generate an initial training sample set. The machine learning model is trained based on the initial training sample set to obtain the dimming decision model.

[0023] In some embodiments, multiple sets of environmental feature information are obtained, the expected illuminance of the dimming amount combination is calculated using the influence coefficient matrix and the natural light illuminance component, the environmental feature information is used as sample input, and the target dimming amount combination that meets the target illuminance and has the lowest energy consumption is used as sample label to generate an initial training sample set, and a machine learning model is trained based on the initial training sample set to obtain the dimming decision model. This can be achieved by the following steps: Multiple sets of personnel distribution characteristics, time period characteristics, and natural lighting characteristics were obtained respectively. Each set of characteristics was normalized and spliced ​​to obtain the corresponding environmental characteristic information and natural light illuminance components. For each set of environmental feature information, the expected illuminance under multiple dimming combinations is calculated using the influence coefficient matrix and the corresponding natural light illuminance component. From the dimming combination that satisfies the target illuminance corresponding to each illuminance sampling point, the target dimming combination with the lowest dimming energy consumption is determined. Each set of environmental feature information is associated with the corresponding target dimming amount combination to form a training sample. The sample input of the training sample is the environmental feature information, and the sample label is the target dimming amount combination. Multiple sets of training samples are collected to obtain the initial training sample set.

[0024] The dimming decision model can be obtained by training a machine learning model based on the initial training sample set using the following steps: The initial training sample set is divided into a training set and a validation set; Using the training set, with sample inputs as model inputs and sample labels as supervised outputs, the model parameters of the machine learning model are iteratively updated through a loss function; Training stops when the loss function value corresponding to the validation set meets the preset convergence condition or reaches the preset number of iterations, thus obtaining the dimming decision model.

[0025] refer to Figure 2 As shown in the figure, this is a schematic flowchart illustrating the process of determining a target dimming amount combination according to some embodiments of this application. The process of determining the target dimming amount combination includes: In step 1021, the target illuminance corresponding to each illuminance sampling point is obtained; In step 1022, candidate dimming combinations that enable the expected illuminance at each illuminance sampling point to reach the target illuminance are selected from the different dimming levels of each downlight. In step 1023, the dimming energy consumption corresponding to each candidate dimming quantity combination is determined, and the candidate dimming quantity combination with the lowest dimming energy consumption is taken as the target dimming quantity combination.

[0026] Regarding the sub-step of calculating the expected illuminance of each downlight at each illuminance sampling point under different dimming levels based on the influence coefficient matrix, in this embodiment, the dimming level of each downlight can be pre-discretely set according to a preset step size within its adjustable range. The value of the step size can be set by balancing the calculation accuracy and computational cost, and here it can be set to 10% of the full range. The dimming level of each downlight is combined one by one to form multiple dimming level combinations. For each dimming level combination, the expected illuminance of each illuminance sampling point under the dimming level combination is output by multiplying the influence coefficient matrix with the dimming level vector corresponding to the dimming level combination and superimposing the natural light illuminance component of the space. The natural light illuminance component can be collected in real time by illuminance sampling points arranged by the window, or it can be obtained by querying a preset natural light illuminance estimation table according to the time period.

[0027] Regarding the sub-step of optimizing the dimming amount of each downlight based on the expected illuminance at each illuminance sampling point, the specific implementation involves first querying the preset standard illuminance value based on the functional type of each area within the space to determine the target illuminance corresponding to each illuminance sampling point. The target illuminance can be set according to the national standard for architectural lighting design, mapped by area function, such as 300 lux for office areas and 100 lux for corridor areas. The target illuminance can be dynamically adjusted based on the presence of people in each area, with the target illuminance being lowered for areas where people have left to free up energy-saving space. Subsequently, all dimming amount combinations are iterated, and dimming amount combinations where the expected illuminance at each illuminance sampling point is not lower than the corresponding target illuminance are retained as candidate dimming amount combinations. Then, for each candidate dimming amount combination, its dimming energy consumption is determined based on the preset downlight power model. The power model expresses the downlight power as a power function relationship between rated power and dimming amount. The power exponent is obtained through actual measurement and calibration of downlights in use. For LED downlights, the power and dimming amount are approximately linearly related, and the power exponent can be a value close to one. The dimming energy consumption of each candidate dimming amount combination is compared, and the candidate dimming amount combination with the lowest dimming energy consumption is output as the target dimming amount combination for each downlight. Preferably, when the number of downlights is large, resulting in a large scale of dimming amount combinations, the above screening and comparison process can be transformed into an optimization problem with the objective of minimizing dimming energy consumption and the constraint that the expected illuminance at each illuminance sampling point is not lower than the target illuminance. The optimal dimming amount can be directly solved using a quadratic programming algorithm. In other embodiments, integer programming or heuristic search algorithms can also be used to solve the optimization problem. This application does not limit this.

[0028] After obtaining the target dimming combination, the environmental feature information of the space is used as the sample input, and the target dimming combination is used as the sample label. The two are associated to form a training sample. In this embodiment, the environmental feature information consists of three parts: personnel distribution features, time period features, and natural lighting features. These respectively characterize the location of personnel in the space, the current time period, and the intensity of natural lighting. After normalization, they are concatenated into a feature vector. By changing the personnel distribution, time period, and natural lighting conditions, the above-mentioned expected illuminance calculation and target dimming combination determination process is repeated to generate training samples covering multiple lighting scenarios in batches, forming an initial training sample set. For example, the input of a training sample is the scene features of a weekday morning, where personnel are concentrated in the east-facing workstations and the natural light illuminance by the window is 200 lux. Its sample label is the target dimming combination obtained by optimization in this scene.

[0029] When training the machine learning model based on the initial training sample set, this embodiment uses a backpropagation neural network. The number of nodes in the input layer is consistent with the feature dimension of the environmental feature information, the number of nodes in the output layer is consistent with the number of downlights, and one or two hidden layers are set. The activation function of the hidden layers can be a modified linear unit. During training, the initial training sample set is divided into a training set and a validation set. A gradient descent optimization algorithm is used to iteratively update the network weights with mean squared error as the loss function. The learning rate can be in the range of 0.1% to 0.01%, and a smaller value is used when the sample size is large to ensure convergence stability. When the loss function value on the validation set no longer decreases for several consecutive rounds or reaches the preset upper limit of the number of iterations, the training is considered to have converged, and the dimming decision model is obtained. This network structure supports continuing training with the trained weights as the starting point, combined with the updated sample set. In subsequent model updates, the above loss function and validation stopping condition are used to update the network weights. The mixing ratio of new and old training samples in the updated sample set changes the proportion of corresponding working condition samples in the training process, so that the matrix correction magnitude enters the model parameter update process through the sample composition of the updated sample set.

[0030] In step 103, the current environmental feature information is input into the dimming decision model, the target dimming amount of each downlight is output, and the dimming of each downlight is controlled.

[0031] In some embodiments, reference Figure 3 As shown in the figure, this is a schematic diagram of the dimming control process according to some embodiments of this application. Inputting current environmental characteristic information into the dimming decision model and outputting the target dimming amount for each downlight, and controlling the dimming of each downlight, can be achieved through the following steps: Real-time data collection of current personnel distribution characteristics, time period characteristics, and natural lighting characteristics is used, and the current environmental feature information is obtained by adopting the same feature composition and normalization method as in the model training phase. The current environmental characteristic information is input into the dimming decision model, and the target dimming amount combination consisting of the target dimming amount of each downlight is output. A dimming control command is generated based on the target dimming amount combination, and the dimming control command is sent to each downlight to control each downlight to perform dimming.

[0032] In specific implementation, personnel sensors deployed within the space can collect the presence status of personnel by area to form personnel distribution characteristics. The system clock reads the current time and maps it to time period characteristics. Illuminance sampling points placed near windows collect natural light illuminance to form natural lighting characteristics. The collection of environmental feature information can be triggered according to a preset period or when personnel are detected entering or leaving the space. The setting of the trigger period can be balanced between response timeliness and system overhead. It should be noted that the environmental feature information collected in real time needs to be processed using the same feature composition and normalization method as in the training phase before being input into the dimming decision model to ensure the consistency of the model input.

[0033] After obtaining the real-time collected environmental feature information, it is concatenated into a feature vector and input into the dimming decision model. After forward calculation by the dimming decision model, the target dimming amount of each downlight is output. The structure and training process of the dimming decision model have been disclosed in the above embodiments and will not be repeated here.

[0034] In this embodiment, the target dimming amount of each downlight can be mapped to a dimming control command recognizable by the dimming drive interface. This command is then sent to each downlight via the dimming drive interface, driving each downlight to adjust its output power to the level corresponding to the target dimming amount. The dimming drive interface can adopt a digital addressable lighting interface protocol, a zero-to-ten-volt analog dimming signal, or a pulse width modulation method. Preferably, before generating the dimming control command, the target dimming amount can be processed for comfort. On the one hand, the rate of change of dimming amount between two adjacent dimming controls is limited to a preset range to avoid visual discomfort caused by sudden changes in illuminance. On the other hand, the target dimming amount is mapped using a gamma curve before being output, which conforms to the logarithmic perception characteristics of human eyes for changes in brightness and improves the linearity of dimming perception in the low-brightness range. After each downlight dims according to the target dimming amount, the illuminance at each illuminance sampling point in the space changes accordingly. Collecting the actual illuminance at this time can provide a data basis for subsequent illuminance deviation correction.

[0035] In step 104, based on the total illuminance of each illuminance sampling point before dimming, the current dimming amount of each downlight, and the actual illuminance after dimming, the natural light illuminance component, expected illuminance, and illuminance deviation are determined using the influence coefficient matrix, and the influence coefficient matrix is ​​corrected based on the changes in dimming amount and illuminance.

[0036] In some embodiments, the natural light illuminance component, expected illuminance, and illuminance deviation are determined using the influence coefficient matrix based on the total illuminance at each illuminance sampling point before dimming, the current dimming level of each downlight, and the actual illuminance after dimming, and the changes in dimming level and illuminance are used to correct the influence coefficient matrix. This can be achieved through the following steps: Before dimming is performed, record the total illuminance at each illuminance sampling point and the current dimming level of each downlight, and collect the actual illuminance when the illuminance stabilizes after dimming is performed. Subtract the product of the influence coefficient matrix and the vector corresponding to the current dimming amount from the total illuminance to obtain the natural light illuminance component. Use the influence coefficient matrix, the target dimming amount, and the natural light illuminance component to calculate the expected illuminance and the illuminance deviation. When the illuminance deviation exceeds the preset deviation threshold, the influence coefficient matrix is ​​corrected by using the changes in dimming amount and illuminance before and after multiple dimming operations as observations, and by employing a recursive least squares algorithm or a batch least squares algorithm.

[0037] In step 105, training samples are regenerated using the corrected influence coefficient matrix. The mixing ratio of the old and new training samples is determined according to the matrix correction magnitude to form an updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point to continue training using the updated sample set, thereby obtaining an updated dimming decision model for subsequent dimming operations.

[0038] In some embodiments, training samples are regenerated using the corrected influence coefficient matrix, and the mixing ratio of the old and new training samples is determined according to the matrix correction magnitude to form an updated sample set. When the cumulative number of times the illuminance deviation exceeds a preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point, and the updated sample set is used to continue training to obtain an updated dimming decision model for subsequent dimming operations. Specifically, this includes: The process of re-calculating the expected illuminance and determining the target dimming amount combination is re-executed using the corrected influence coefficient matrix to generate new training samples. The matrix correction magnitude is determined based on the difference before and after the matrix correction. The mixing ratio of the new training samples and the original training samples is determined according to the matrix correction magnitude. The larger the matrix correction magnitude, the higher the mixing ratio of the new training samples, thus obtaining the updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the training starting point, and the updated sample set is used to continue training to obtain the updated dimming decision model, which is then used for subsequent dimming operations.

[0039] In another embodiment, when the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the updated sample set can be used to fully retrain the machine learning model, and the model obtained from the full retraining can be used for subsequent dimming operations.

[0040] In practice, after each downlight is dimmed to the target dimming level and the illuminance stabilizes, the actual illuminance at each illuminance sampling point is collected to avoid the interference of instantaneous fluctuations during the dimming transition process in the collected results. At the same time, the natural light illuminance component of the space is estimated. In this embodiment, the current total illuminance at each illuminance sampling point and the current dimming level of each downlight can be recorded before dimming is executed. The difference between the current total illuminance and the product of the influence coefficient matrix and the vector corresponding to the current dimming level is the natural light illuminance component. Since dimming execution and illuminance collection are completed within a short period of time, under the condition that the short-term change in natural light is not significant, the natural light illuminance component during this period can be regarded as an approximate invariant. It should be noted that the above estimation method reuses the influence coefficient matrix to separate the contribution of natural light from the total illuminance. The downlights do not need to be turned off throughout the process, nor does it interrupt the normal lighting use of the space.

[0041] After obtaining the natural light illuminance component, the expected illuminance of each illuminance sampling point is calculated using the influence coefficient matrix and the dimming amount executed by each downlight. That is, the influence coefficient matrix is ​​multiplied by the corresponding vector of the dimming amount executed and the natural light illuminance component is superimposed. Then, the actual illuminance of each illuminance sampling point is subtracted from the corresponding expected illuminance to obtain the illuminance deviation of each illuminance sampling point. The illuminance deviation is used as an observation of the degree of mismatch between the influence coefficient matrix and the current lighting response.

[0042] Subsequently, when correcting the influence coefficient matrix based on the illuminance deviation, specifically, it can first be determined whether the illuminance deviation of each illuminance sampling point exceeds a preset deviation threshold. The value of the deviation threshold can be set according to the accuracy level of illuminance sampling and the allowable illuminance control error. For example, it can be 10% of the target illuminance. When the accuracy level is high, it can be adjusted to be smaller to improve the correction sensitivity. Illuminance sampling points whose illuminance deviation does not exceed the deviation threshold are not processed to avoid frequent changes in the matrix caused by sensor noise. For illuminance sampling points that exceed the deviation threshold, the influence coefficient matrix is ​​corrected online using a recursive least squares algorithm. The row corresponding to the illuminance sampling point in the influence coefficient matrix is ​​updated recursively based on the changes in dimming amount and illuminance before and after the dimming operation. The observations are accumulated and continuously corrected with each dimming operation. Preferably, a forgetting factor is introduced in the recursive process to give higher weight to recent observations. The forgetting factor can be in the range of 0.95 to 1. The faster the lamp ages, the smaller the value should be, so that the correction result can keep up with the light decay of the lamp in a timely manner. In other embodiments, the changes in dimming amount and illuminance after multiple dimming operations can be accumulated and then batch-resolved with least squares. This application does not limit this.

[0043] Finally, when updating the dimming decision model using the corrected influence coefficient matrix, the process of calculating the expected illuminance and determining the target dimming amount combination is repeated using the corrected influence coefficient matrix to regenerate training samples. The regenerated training samples are then mixed with the original training samples according to the proportion determined by the matrix correction magnitude to obtain an updated sample set. The larger the matrix correction magnitude, the higher the proportion of the regenerated training samples. The updated sample set serves as the input for subsequent training, and its sample composition influences the degree of correction of the illuminance contribution relationship in the model parameter update process. The retraining is triggered when the cumulative number of times the illuminance deviation exceeds the deviation threshold reaches a preset number, to avoid triggering training with a single deviation. After triggering, starting from the existing weights of the dimming decision model, the loss function and verification stopping condition from step 102 are used to continue training using the updated sample set to obtain the updated dimming decision model. The updated dimming decision model is used for subsequent dimming operations, allowing physical response drift to be sequentially incorporated into training data composition, model parameter updates, and subsequent real-world lamp control, forming a closed loop of physical response correction and model adaptive updates.

[0044] Furthermore, in another aspect of this application, in some embodiments, this application provides a machine learning-based intelligent decision-making system for energy-saving dimming of downlights, referencing... Figure 4 The system 200 includes a response modeling module 201, an initial training module 202, a decision execution module 203, a drift observation module 204, and a model update module 205. Each module is described below: The response modeling module 201 is used to change the dimming amount of each downlight while keeping the dimming amount of the remaining downlights constant, and to construct an influence coefficient matrix based on the changes in dimming amount and the changes in illuminance at each illuminance sampling point. The initial training module 202 is used to acquire multiple sets of environmental feature information, calculate the expected illuminance of the dimming amount combination using the influence coefficient matrix and the natural light illuminance component, use the environmental feature information as sample input, and the target dimming amount combination that meets the target illuminance and has the lowest energy consumption as sample label to generate an initial training sample set and train a machine learning model based on the initial training sample set to obtain a dimming decision model. The decision execution module 203 is used to input the current environmental feature information into the dimming decision model, output the target dimming amount of each downlight, and control the dimming of each downlight; The drift observation module 204 is used to determine the natural light illuminance component, expected illuminance and illuminance deviation based on the total illuminance of each illuminance sampling point before dimming execution, the current dimming amount of each downlight and the actual illuminance after execution, using the influence coefficient matrix, and to correct the influence coefficient matrix based on the changes in dimming amount and illuminance. The model update module 205 is used to regenerate training samples using the corrected influence coefficient matrix, determine the mixing ratio of the old and new training samples according to the matrix correction magnitude, and form an updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point, and the updated sample set is used to continue training to obtain the updated dimming decision model for subsequent dimming operation.

[0045] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described machine learning-based intelligent decision-making method for downlight energy-saving dimming.

[0046] In some embodiments, reference Figure 5 This figure is an internal structural diagram of a computer device implementing a machine learning-based intelligent decision-making method for downlight energy-saving dimming, according to some embodiments of this application. The machine learning-based intelligent decision-making method for downlight energy-saving dimming in the above embodiments can be implemented through... Figure 5 The computer device shown is used to implement this, and the computer device 300 includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0047] The processor 301 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the machine learning-based intelligent decision-making method for energy-saving dimming of downlights in this application.

[0048] The communication bus 302 is used to transmit information between the aforementioned components.

[0049] Memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 303 may exist independently and be connected to processor 301 via communication bus 302. Memory 303 may also be integrated with processor 301.

[0050] The memory 303 stores program code for executing the solution of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the machine learning-based intelligent decision-making method for downlight energy saving and dimming can be implemented by the processor 301 and one or more software modules in the program code in the memory 303.

[0051] Communication interface 304 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0052] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0053] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device may be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0054] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent decision-making method for energy-saving dimming of downlights based on machine learning.

[0055] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0056] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A machine learning-based intelligent decision-making method for energy-saving dimming of downlights, characterized in that, Includes the following steps: The dimming level of each downlight is changed while keeping the dimming level of the remaining downlights constant. An influence coefficient matrix is ​​constructed based on the changes in dimming level and the changes in illuminance at each illuminance sampling point. Multiple sets of environmental feature information are obtained, and the expected illuminance of the dimming amount combination is calculated using the influence coefficient matrix and the natural light illuminance component. The environmental feature information is used as the sample input, and the target dimming amount combination that meets the target illuminance and has the lowest energy consumption is used as the sample label. An initial training sample set is generated, and a machine learning model is trained based on the initial training sample set to obtain a dimming decision model. The current environmental characteristics information is input into the dimming decision model, which outputs the target dimming amount for each downlight and controls the dimming of each downlight. Based on the total illuminance of each illuminance sampling point before dimming, the current dimming amount of each downlight, and the actual illuminance after dimming, the natural light illuminance component, expected illuminance, and illuminance deviation are determined using the influence coefficient matrix. The influence coefficient matrix is ​​then corrected based on the changes in dimming amount and illuminance. The training samples are regenerated using the corrected influence coefficient matrix. The mixing ratio of the old and new training samples is determined according to the matrix correction magnitude to form an updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point to continue training using the updated sample set, thereby obtaining an updated dimming decision model for subsequent dimming operations.

2. The method as described in claim 1, characterized in that, The dimming level of each downlight is changed while keeping the dimming level of the remaining downlights constant. An influence coefficient matrix is ​​constructed based on the changes in dimming level and the changes in illuminance at each illuminance sampling point. Specifically, this includes: Select a single downlight as the disturbance object according to a preset order, keep the dimming amount of the other downlights unchanged, and make the dimming amount of the disturbance object jump from the current value to another dimming amount within a short time window, and simultaneously collect the illuminance values ​​of each illuminance sampling point before and after the jump. Determine the illuminance change at each illuminance sampling point before and after the jump, divide the illuminance change at each illuminance sampling point by the dimming amplitude of the dimming, and use the illuminance influence coefficient corresponding to the disturbance object. Go through each downlight one by one, and arrange the illuminance influence coefficients according to the correspondence between the illuminance sampling points and the downlights to obtain the influence coefficient matrix.

3. The method as described in claim 1, characterized in that, Obtaining multiple sets of environmental feature information, calculating the expected illuminance of the dimming combination using the influence coefficient matrix and natural light illuminance components, and generating an initial training sample set using the environmental feature information as sample input and the target dimming combination that satisfies the target illuminance and has the lowest energy consumption as sample labels, specifically includes: Multiple sets of personnel distribution characteristics, time period characteristics, and natural lighting characteristics were obtained respectively. Each set of characteristics was normalized and spliced ​​to obtain the corresponding environmental characteristic information and natural light illuminance components. For each set of environmental feature information, the expected illuminance under multiple dimming combinations is calculated using the influence coefficient matrix and the corresponding natural light illuminance component. From the dimming combination that satisfies the target illuminance corresponding to each illuminance sampling point, the target dimming combination with the lowest dimming energy consumption is determined. Each set of environmental feature information is associated with the corresponding target dimming amount combination to form a training sample. The sample input of the training sample is the environmental feature information, and the sample label is the target dimming amount combination. Multiple sets of training samples are collected to obtain the initial training sample set.

4. The method as described in claim 1, characterized in that, The machine learning model trained based on the initial training sample set yields the dimming decision model, which specifically includes: The initial training sample set is divided into a training set and a validation set; Using the training set, with sample inputs as model inputs and sample labels as supervised outputs, the model parameters of the machine learning model are iteratively updated through a loss function; Training stops when the loss function value corresponding to the validation set meets the preset convergence condition or reaches the preset number of iterations, thus obtaining the dimming decision model.

5. The method as described in claim 1, characterized in that, Inputting current environmental characteristic information into the dimming decision model, outputting the target dimming amount for each downlight, and controlling the dimming of each downlight specifically includes: Real-time data collection of current personnel distribution characteristics, time period characteristics, and natural lighting characteristics is used, and the current environmental feature information is obtained by adopting the same feature composition and normalization method as in the model training phase. The current environmental characteristic information is input into the dimming decision model, and the target dimming amount combination consisting of the target dimming amount of each downlight is output. A dimming control command is generated based on the target dimming amount combination, and the dimming control command is sent to each downlight to control each downlight to perform dimming.

6. The method as described in claim 1, characterized in that, Based on the total illuminance at each illuminance sampling point before dimming, the current dimming level of each downlight, and the actual illuminance after dimming, the influence coefficient matrix is ​​used to determine the natural light illuminance component, expected illuminance, and illuminance deviation. The influence coefficient matrix is ​​then corrected based on changes in dimming level and illuminance. Specifically, this includes: Before dimming is performed, record the total illuminance at each illuminance sampling point and the current dimming level of each downlight, and collect the actual illuminance when the illuminance stabilizes after dimming is performed. Subtract the product of the influence coefficient matrix and the vector corresponding to the current dimming amount from the total illuminance to obtain the natural light illuminance component. Use the influence coefficient matrix, the target dimming amount, and the natural light illuminance component to calculate the expected illuminance and the illuminance deviation. When the illuminance deviation exceeds the preset deviation threshold, the influence coefficient matrix is ​​corrected by using the changes in dimming amount and illuminance before and after multiple dimming operations as observations, and by employing a recursive least squares algorithm or a batch least squares algorithm.

7. The method as described in claim 1, characterized in that, The training samples are regenerated using the corrected influence coefficient matrix. The mixing ratio of the old and new training samples is determined according to the matrix correction magnitude to form an updated sample set. When the cumulative number of times the illuminance deviation exceeds a preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point to continue training using the updated sample set, resulting in an updated dimming decision model for subsequent dimming operations. Specifically, this includes: The process of re-calculating the expected illuminance and determining the target dimming amount combination is re-executed using the corrected influence coefficient matrix to generate new training samples. The matrix correction magnitude is determined based on the difference before and after the matrix correction. The mixing ratio of the new training samples and the original training samples is determined according to the matrix correction magnitude. The larger the matrix correction magnitude, the higher the mixing ratio of the new training samples, thus obtaining the updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the training starting point, and the updated sample set is used to continue training to obtain the updated dimming decision model, which is then used for subsequent dimming operations.

8. A machine learning-based intelligent decision-making system for energy-saving dimming of downlights, characterized in that, include: The response modeling module is used to change the dimming level of each downlight while keeping the dimming level of the remaining downlights constant, and to construct an influence coefficient matrix based on the changes in dimming level and the changes in illuminance at each illuminance sampling point. The initial training module is used to acquire multiple sets of environmental feature information, calculate the expected illuminance of the dimming amount combination using the influence coefficient matrix and the natural light illuminance component, use the environmental feature information as sample input, and the target dimming amount combination that meets the target illuminance and has the lowest energy consumption as sample label to generate an initial training sample set and train a machine learning model based on the initial training sample set to obtain a dimming decision model. The decision execution module is used to input the current environmental feature information into the dimming decision model, output the target dimming amount of each downlight, and control the dimming of each downlight. The drift observation module is used to determine the natural light illuminance component, expected illuminance, and illuminance deviation based on the total illuminance of each illuminance sampling point before dimming, the current dimming amount of each downlight, and the actual illuminance after dimming, using the influence coefficient matrix. The influence coefficient matrix is ​​then corrected based on the changes in dimming amount and illuminance. The model update module is used to regenerate training samples using the corrected influence coefficient matrix, determine the mixing ratio of the old and new training samples according to the matrix correction magnitude, and form an updated sample set. When the cumulative number of times the illuminance deviation exceeds the preset deviation threshold reaches a preset number, the existing weights of the dimming decision model are used as the starting point, and the updated sample set is used to continue training to obtain the updated dimming decision model for subsequent dimming operation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent decision-making method for energy-saving dimming of downlights based on machine learning, as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent decision-making method for energy-saving dimming of downlights based on machine learning as described in any one of claims 1 to 7.

Images