A method and system for coordinated control of shading and lighting based on solar thermal load prediction
By constructing a cascaded lighting demand and heat gain prediction model, and a shading and lighting linkage control system based on light and heat load prediction, the problem of the disconnect between the coupling relationship of shading, lighting and air conditioning systems in buildings is solved, realizing forward-looking prediction and energy consumption optimization, and improving the building's comprehensive energy consumption management and light and heat comfort.
Patent Information
- Application Number
- CN202611473301.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-21
- Publication Date
- 2026-11-17
AI Technical Summary
In the existing joint control of building shading, lighting and air conditioning systems, the overall energy consumption optimization effect is limited due to the fragmented coupling relationship among the three, lack of forward-looking prediction, and lack of closed-loop feedback between lighting heat generation and heat load.
By deploying a sensor network to collect environmental and equipment operation data, a cascaded lighting demand prediction sub-model and a thermal gain prediction sub-model are constructed. Based on the prediction of light and heat load, the comprehensive energy consumption is calculated, and the optimal control strategy, including curtain opening degree, target illuminance and temperature setpoint, is determined to achieve light and heat coupling linkage and forward-looking control.
It achieves photothermal coupling and linkage, possesses forward-looking predictive control and comprehensive energy consumption optimization, and improves the photothermal comfort and energy management efficiency of buildings.
Smart Images

Figure CN123113598A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of control technology, and more specifically, to a method and system for shading and lighting linkage control based on solar thermal load prediction. Background Technology
[0002] In the field of intelligent building control, the coordinated operation of shading systems, lighting systems, and air conditioning systems plays a crucial role in reducing overall building energy consumption and improving indoor light and heat comfort. These three systems collectively influence indoor environmental quality and equipment energy consumption by regulating natural lighting and solar radiation heat gain, supplementing artificial lighting, and maintaining indoor temperature, respectively. However, in existing technologies, these three systems typically operate independently or only undergo simple linkage control. This results in shading devices, while blocking solar radiation to reduce air conditioning load, also reducing natural lighting, thus forcing an increase in artificial lighting energy consumption. This passive transfer of energy consumption between different energy forms means that energy-saving measures for a single system may actually increase overall building energy consumption. Furthermore, existing control methods often employ reactive control strategies, adjusting based on current environmental conditions without the ability to predict future trends in light and heat load. Lagging adjustment methods are difficult to implement effectively. Currently, there are several issues with control optimization. Furthermore, the heat generated by lighting equipment constitutes a significant portion of the indoor heat load. However, in existing technologies, the dynamic impact of lighting heat generation on air conditioning load is often ignored or roughly estimated using fixed coefficients, leading to insufficient load prediction accuracy. Moreover, the physical mechanisms of light environment evaluation indicators and thermal environment evaluation indicators differ, making it difficult to simultaneously consider both visual and thermal comfort requirements with a single objective optimization. More critically, while some existing joint control schemes involve the coordination of shading, lighting, and air conditioning, they employ a hierarchical or sequential optimization architecture. This involves first optimizing shading and lighting to reduce thermal disturbance, and then optimizing the air conditioning system based on the corrected thermal disturbance value. This hierarchical architecture severs the real-time physical coupling feedback link between lighting power demand and thermal gain, preventing adjustments to lighting power from reflecting their impact on the heat load in real time at the same optimization level, resulting in decision-making bias. Summary of the Invention
[0003] This application aims to provide a shading and lighting linkage control method and system based on solar and thermal load prediction, which aims to solve the technical problems that limit the overall energy consumption optimization effect in the existing joint control of building shading, lighting and air conditioning due to the disconnect of the coupling relationship among the three, lack of forward-looking prediction and lack of closed-loop feedback between lighting heat generation and heat load.
[0004] On one hand, the present invention provides a shading and lighting linkage control method based on solar thermal load prediction, comprising: Environmental data and equipment operation data are collected through an indoor and outdoor sensor network, including indoor illuminance, outdoor illuminance, solar radiation intensity, indoor temperature, curtain opening, lighting power, and air conditioning load data. Construct and pre-train a solar and thermal load prediction model; the solar and thermal load prediction model includes a cascaded lighting demand prediction sub-model and a thermal gain prediction sub-model. The environmental data and equipment operation data are input into the lighting demand prediction sub-model to obtain the predicted value of lighting power demand. The input features of the thermal gain prediction sub-model are constructed based on the predicted lighting power demand value. Together with the environmental data and equipment operation data, the input of the thermal gain prediction sub-model is constructed to obtain the indoor thermal gain prediction value. The comprehensive energy consumption is calculated based on the predicted lighting power demand and the predicted indoor thermal gain. Under the conditions of satisfying the preset illuminance constraints, preset temperature constraints and preset glare constraints, the optimal control strategy is determined with the goal of minimizing the comprehensive energy consumption. The optimal control strategy includes the curtain opening degree, target illuminance and temperature setpoint.
[0005] Furthermore, before inputting the environmental data and equipment operation data into the lighting demand prediction sub-model, the following steps are also included: The environmental data and equipment operation data are preprocessed, including sliding window filtering, outlier detection and removal based on the 3σ criterion, forward filling and masking of missing values, and / or minimum-maximum normalization.
[0006] Furthermore, the lighting demand prediction sub-model is set as a long short-term memory network model; The input features of the lighting demand prediction sub-model include normalized outdoor illuminance, normalized solar radiation intensity, normalized curtain opening, normalized indoor illuminance, cyclic time-coded sine component, cyclic time-coded cosine component, and input feature validity mask. The output of the lighting demand prediction sub-model is the normalized value of the lighting power required to maintain the target illuminance at the next time step.
[0007] Furthermore, the thermal gain prediction sub-model is set as a long short-term memory network model; The input features of the thermal gain prediction sub-model include normalized values of solar radiation intensity, normalized values of curtain opening, normalized values of outdoor temperature, normalized values of indoor temperature, normalized values of lighting power, and a mask for the validity of input features. The output of the thermal gain prediction sub-model is the predicted value of the total indoor thermal gain at the next time step.
[0008] Furthermore, the training loss function used by the thermal gain prediction sub-model during training includes a physical consistency regularization term, which is used to apply a positive penalty when the indoor thermal gain prediction value decreases as the solar radiation intensity increases.
[0009] Furthermore, the comprehensive energy consumption is the sum of the lighting power demand and the air conditioning power demand. The air conditioning power demand is calculated based on the indoor heat gain prediction and the air conditioning energy efficiency ratio. The air conditioning energy efficiency ratio is determined based on the measured outdoor temperature and the unit's partial load rate.
[0010] Furthermore, the preset glare constraint is evaluated using a glare risk index proxy model; the glare risk index is calculated based on the equivalent light transmittance of the roller blind, the measured value of solar radiation intensity, the measured value of outdoor illuminance, and the value of lighting power demand; when all candidate control strategies fail to meet the preset glare constraint, a multi-level constraint relaxation mechanism is executed sequentially.
[0011] Furthermore, determining the optimal control strategy with the goal of minimizing the overall energy consumption includes: Multiple candidate control strategies are generated, each candidate control strategy including a combination of candidate values for curtain opening, candidate values for target illuminance, and candidate values for temperature setpoint; Calculate the overall energy consumption for each candidate control strategy; Among the candidate control strategies that satisfy the preset illuminance constraint, the preset temperature constraint, and the preset glare constraint, the candidate control strategy with the minimum overall energy consumption is selected as the optimal control strategy.
[0012] Furthermore, determining the optimal control strategy with the goal of minimizing the overall energy consumption also includes: A penalty term for changes in control action is introduced into the objective function; the penalty term for changes in control action includes the square penalty for changes in curtain opening, the square penalty for changes in temperature setpoint, and the square penalty for changes in target illuminance; the optimal control strategy is the candidate control strategy that minimizes the sum of the overall energy consumption and the penalty term for changes in control action.
[0013] On the other hand, the present invention provides a shading and lighting linkage control system based on solar thermal load prediction, comprising: The data acquisition module is used to collect environmental data and equipment operation data through a sensor network deployed indoors and outdoors. The environmental data and equipment operation data include indoor illuminance, outdoor illuminance, solar radiation intensity, indoor temperature, curtain opening degree, lighting power and air conditioning load data. The lighting prediction module is used to build and pre-train a solar and thermal load prediction model. The solar and thermal load prediction model includes a cascaded lighting demand prediction sub-model and a thermal gain prediction sub-model. The environmental data and equipment operation data are input into the lighting demand prediction sub-model to obtain the predicted value of lighting power demand. The thermal gain prediction module is used to construct the input features of the thermal gain prediction sub-model based on the predicted lighting power demand value, and to construct the input of the thermal gain prediction sub-model together with the environmental data and equipment operation data to obtain the indoor thermal gain prediction value. The linkage control module is used to calculate the comprehensive energy consumption based on the predicted lighting power demand and the predicted indoor heat gain. Under the conditions of satisfying the preset illuminance constraints, preset temperature constraints and preset glare constraints, the optimal control strategy is determined with the goal of minimizing the comprehensive energy consumption. The optimal control strategy includes the curtain opening degree, target illuminance and temperature setpoint.
[0014] This application provides a method and system for coordinated control of shading lighting based on solar and thermal load prediction. It collects environmental and equipment operation data through a sensor network; constructs cascaded lighting demand prediction sub-models and thermal gain prediction sub-models; inputs the collected data into the lighting demand prediction sub-model to obtain a predicted lighting power demand value; uses this predicted value as an input feature of the thermal gain prediction sub-model, inputting it along with the collected data to obtain an indoor thermal gain prediction value; calculates the comprehensive energy consumption based on the predicted lighting power demand value and the indoor thermal gain prediction value; and determines the optimal control strategy, including curtain opening degree, target illuminance, and temperature setpoints, with the goal of minimizing comprehensive energy consumption, while satisfying illuminance, temperature, and glare constraints. This invention has the advantages of solar and thermal coupling, forward-looking predictive control, and comprehensive energy consumption optimization. Attached Figure Description
[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart of a shading and lighting linkage control method based on solar thermal load prediction proposed in an embodiment of this application; Figure 2 This is a schematic diagram of a shading and lighting linkage control system based on solar thermal load prediction, according to an embodiment of this application. Among them, 201-data acquisition module, 202-lighting prediction module, 203-thermal gain prediction module, and 204-linkage control module. Detailed Implementation
[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0018] like Figure 1 As shown in the figure, this application provides a shading and lighting linkage control method based on solar thermal load prediction, including: S101. Collect environmental data and equipment operation data through an indoor and outdoor sensor network. The environmental data and equipment operation data include indoor illuminance, outdoor illuminance, solar radiation intensity, indoor temperature, curtain opening, lighting power and air conditioning load data. In this embodiment of the invention, before inputting environmental data and equipment operation data into the lighting demand prediction sub-model, the method further includes: Environmental data and equipment operation data are preprocessed, including sliding window filtering, outlier detection and removal based on the 3σ criterion, forward filling and masking of missing values, and / or minimum-maximum normalization.
[0019] For example, raw environmental and equipment operating status data are collected by a sensor network deployed in specific indoor and outdoor locations. Then, the raw data is processed by sliding window filtering, outlier detection and missing value forward filling and masking based on the 3σ criterion, and minimum-maximum normalization. Finally, a standardized time series dataset that can be directly input into a long short-term memory network model is formed, providing time-aligned and formatted input time series sequences for subsequent data-driven cascaded prediction models.
[0020] To achieve comprehensive perception of indoor and outdoor light and heat environments, this embodiment deploys a multi-sensor network in a single-story office area of the target building. It is assumed that the standard floor area of the single-story office area is 1200 square meters, and the south-facing window area accounts for 40%.
[0021] The indoor illuminance sensors are deployed at a height of 0.75 meters above the ground on the work surface, with one measuring point for every 50 square meters, for a total of 24 indoor illuminance sensor measuring points. The measurement range of the indoor illuminance sensor is 0 to 2000 lux and the measurement accuracy is ±3% of the reading.
[0022] The outdoor illuminance sensor is deployed at an unobstructed location on the south facade of the building's roof, with a total of two sensor measurement points. The outdoor illuminance sensor has a measurement range of 0 to 100,000 lux and a measurement accuracy of ±5% of the reading.
[0023] The solar radiation sensor uses a total radiation meter and is installed in the same location as the outdoor illuminance sensor, with a total of two solar radiation sensor measuring points deployed. The solar radiation sensor has a measurement range of 0 to 1500 watts per square meter and an accuracy class of 2, with an accuracy of ±5%.
[0024] The indoor temperature sensor is deployed at a height of 1.2 meters above the ground, with one measuring point for every 100 square meters, for a total of 12 indoor temperature sensor measuring points. The indoor temperature sensor has a measurement range of -10 to +50 degrees Celsius and a measurement accuracy of ±0.2 degrees Celsius.
[0025] The curtain opening encoder is coaxially mounted at the tail of each electric roller blind motor, with a total of 48 encoder measurement points. The encoder has a measurement range of 0% to 100% and a measurement accuracy of ±1%.
[0026] The lighting power smart meter is installed in series at the outgoing terminal of each lighting distribution circuit, with a total of 8 lighting power smart meter measuring points deployed. The measuring range of this lighting power smart meter is 0 to 10 kilowatts, the accuracy class is 1, and the accuracy is ±1%.
[0027] The air conditioning cooling and heating meter is installed on the chilled water supply and return main pipe of the air conditioning unit and includes an ultrasonic flow meter and a temperature sensor pair. The flow measurement range of this air conditioning cooling and heating meter is 0 to 50 cubic meters per hour, with a flow measurement accuracy of ±2%; the temperature difference measurement range is 0 to 15 degrees Celsius, with a temperature difference measurement accuracy of ±0.1 degrees Celsius.
[0028] All sensors are connected to the edge acquisition gateway via RS-485 bus or wireless LoRa communication. This embodiment employs the IEEE 1588 precise time synchronization protocol to synchronize the clocks of all sensors in the network, ensuring that the deviation of each sensor's acquisition time is less than ±10 milliseconds. This synchronization accuracy is sufficient to guarantee that the data at each time step in the subsequent input sequence of the Long Short-Term Memory network model physically corresponds to the same moment, avoiding the introduction of false causal relationships due to timing misalignment.
[0029] In this embodiment, the sampling period is set to 60 seconds, meaning that a network-wide synchronous data collection is triggered every whole minute. The sampling period is determined based on the fact that the building thermal inertia time constant is usually on the order of 10 to 30 minutes. According to the Nyquist sampling theorem, the sampling period should be less than half of the thermal inertia time constant (i.e., less than 5 minutes). The 60-second sampling period meets this requirement while also taking into account the overhead of data storage and computing resources.
[0030] At each acquisition moment, the data acquisition module obtains a multi-dimensional observation vector, which sequentially contains the following data: illuminance values from 24 indoor illuminance sensor points, illuminance values from 2 outdoor illuminance sensor points, radiation intensity values from 2 solar radiation sensor points, temperature values from 12 indoor temperature sensor points, opening percentage values from 48 curtain opening encoder points, power values from 8 smart lighting power meters, chilled water flow rate value, and supply and return water temperature difference value. All data are stored in a time-series database according to timestamps; in this embodiment, InfluxDB is used as the time-series database.
[0031] The real-time heat load of the air conditioner is not directly read from the sensors, but is calculated in real time by the edge acquisition gateway based on the chilled water volumetric flow rate and the supply and return water temperature difference. The calculation method is as follows: multiply the specific heat capacity of water at constant pressure (4.18 × 10³ joules per kilogram of temperature) by the density of water (1.0 × 10³ kilograms per cubic meter), and then multiply by the chilled water volumetric flow rate and the supply and return water temperature difference. In a specific example, the chilled water volumetric flow rate was collected as 0.012 cubic meters per second, and the supply and return water temperature difference was 5.2 degrees Celsius. Substituting these values into the calculation, the real-time heat load of the air conditioner was found to be 260,832 watts, approximately equal to 260.8 kilowatts.
[0032] Through the aforementioned sensor deployment, time synchronization, multi-dimensional observation vector construction, and air conditioning heat load calculation, this step achieves a comprehensive perception of the target building's light and heat environment and equipment operating status, providing a physically meaningful and time-aligned original data foundation for subsequent prediction models.
[0033] This embodiment employs sliding window mean filtering to smooth continuous sensor signals. Continuous sensor signals include illuminance sensor signals, radiation sensor signals, and temperature sensor signals. The method processes each sensor channel separately: for a given sensor channel, the values at the current moment and several previous sampling moments are arithmetically averaged, and this average is used as the filtered value at that moment. The number of sampling points involved in the averaging is called the sliding window width.
[0034] The window width was determined through offline calibration. The calibration method was as follows: before formal control operation, data from one typical sunny day and one typical cloudy day were selected as the calibration dataset. For each sensor channel, filtering was performed with window widths of 3, 5, 7, 10, and 15, respectively. The root mean square error (RMSE) of the filtered sequence compared to the manually labeled noise-free reference segment in the original sequence was used as the evaluation metric. The early morning period (during which there is no light and the temperature is stable, and the signal truth value is approximately constant) was selected as the noise-free reference segment. Experimental statistical results showed that the RMSE was smallest when the window width was 5, with a noise suppression rate improvement of approximately 42% compared to a window width of 3. However, compared to a window width of 7, the RMSE actually increased by approximately 8% due to the increased signal delay. Therefore, in this embodiment, the window width was fixed at 5.
[0035] In a specific example, the original sampled values of an indoor illuminance sensor channel at five consecutive sampling times were 523 lux, 541 lux, 512 lux, 536 lux, and 529 lux, respectively. After sliding window mean filtering, the filtered value was 528.2 lux. The filtered value differed from the latest original value of 529 lux by less than 1%, while effectively suppressing the transient jumps at 541 lux and 512 lux.
[0036] By using this moving average filtering process, this step reduces the standard deviation of high-frequency random noise to about 44.7% of the original noise without introducing significant signal delay (the delay is about 2 minutes, which is much smaller than the building thermal inertia time constant). This solves the technical problem of sensor noise contaminating the training data of subsequent long short-term memory network models.
[0037] This embodiment employs a sliding window anomaly detection method based on the 3σ criterion. For each sensor channel, a historical window with a length of 60 sampling points is maintained at the current moment. First, the arithmetic mean of the data within the window is calculated, and then the sample standard deviation of the data within the window is calculated. If the original sampled value at the current moment deviates from the window mean by more than 3 times the standard deviation, the value is determined to be an anomaly. The window length is determined based on the fact that a 60-minute timescale is sufficient to cover the short-term fluctuation patterns of natural variations in solar radiation, while being much smaller than the timescale of weather system evolution.
[0038] For sampling points identified as outliers, this embodiment does not perform interpolation imputation but directly discards the data for that channel at that moment, marking it as a null value on the corresponding data point in the time-series database. For subsequent model inputs, missing values are filled using a forward imputation strategy, i.e., replaced by the most recent non-missing value. Simultaneously, a binary mask feature is added to each sensor channel, with a value of 0 indicating that the data at that moment is an original valid value, and a value of 1 indicating that the data at that moment is an imputed value. This mask feature will serve as an additional input feature for the Long Short-Term Memory network model, enabling the model to perceive the confidence information of the data.
[0039] In a specific example, the current value collected by an outdoor illuminance sensor channel is 47200 lux, with a historical window mean of 45800 lux and a standard deviation of 320 lux. The absolute value of the deviation is 1400 lux, exceeding three times the standard deviation (960 lux), therefore this value is determined to be an outlier. After investigation, it was found that this jump was caused by a communication data frame verification error. The data at this moment was marked as null, and forward padding was performed using the previous valid value of 45780 lux, with the mask bit set to 1.
[0040] Through the outlier detection and forward padding masking mechanism described above, this step eliminates physically invalid data caused by communication failures, while retaining data confidence information through mask features, thus avoiding false correlations introduced by simple interpolation.
[0041] In this embodiment, minimum-maximum normalization is performed on each sensor channel, mapping all features to a closed interval between 0 and 1. For a given sensor channel, the normalized value is calculated as follows: subtract the minimum value of that channel in the entire training set from the original value at that moment, and then divide by the difference between the maximum and minimum values of that channel in the entire training set.
[0042] The maximum and minimum values are determined from the training set data. In this embodiment, during the offline training phase, all historical data is used as the training set, and the extreme values of each channel are statistically analyzed and stored as a normalization parameter configuration file. During the online inference phase, all newly acquired data is normalized using the same offline-determined extreme value. If the system has been running for more than a year and new data exceeding the historical extreme value is detected, the offline recalibration process is triggered.
[0043] In a specific example, the outdoor illuminance channel has a minimum value of 0 lux and a maximum value of 98,500 lux in the entire training set. At a certain moment, the original value was 65,000 lux, which is approximately 0.660 after normalization. The lighting power channel has a minimum value of 0 watts and a maximum value of 8,500 watts in the entire training set. At a certain moment, the original value was 3,200 watts, which is approximately 0.376 after normalization.
[0044] Through the above min-maximum normalization process, all input features are mapped to the same numerical scale, eliminating the influence of dimensional differences on gradient descent, and enabling subsequent long short-term memory network prediction models to update weights with a uniform step size.
[0045] After the aforementioned steps, a normalized feature vector is obtained at each time step, containing the normalized values of all sensor channels and their corresponding mask features, with a total dimension of 99. To construct training samples for the Long Short-Term Memory network, for each time step, the normalized feature vectors of that time step and several consecutive time steps prior are extracted to form a three-dimensional tensor. The first dimension of this tensor is the time step, and the second dimension is the feature.
[0046] In this embodiment, the input sequence length is set to 30, corresponding to a 30-minute historical data window. This value is determined through grid search hyperparameter calibration in subsequent steps, and it can cover the main dynamic processes of the shading motor's response time and indoor temperature changes. The corresponding prediction target is the target value at the next moment; the prediction target of the lighting demand prediction sub-model is the lighting power demand at the next moment, and the prediction target of the thermal gain prediction sub-model is the indoor thermal gain at the next moment.
[0047] The processed, standardized data was organized into a three-dimensional tensor format that is directly usable by the Long Short-Term Memory network model, providing standardized data input for subsequent model training.
[0048] It is worth noting that the specific values of the above parameters are only examples in the embodiments of this application, and can be collected according to the actual situation.
[0049] S102. Construct and pre-train the solar and thermal load prediction model. The solar and thermal load prediction model includes a cascaded lighting demand prediction sub-model and a thermal gain prediction sub-model. Environmental data and equipment operation data are input into the lighting demand prediction sub-model to obtain the predicted value of lighting power demand. In this embodiment of the invention, the lighting demand prediction sub-model is set as a long short-term memory network model; The input features of the lighting demand prediction sub-model include normalized outdoor illuminance, normalized solar radiation intensity, normalized curtain opening, normalized indoor illuminance, cyclic time-coded sine component, cyclic time-coded cosine component, and an input feature validity mask. The input feature validity mask is defined as follows: for data from a certain sensor channel at a certain time, if the original sampled value at that time is determined to be valid and has not been filled, the corresponding mask value for that time is recorded as 0; if the original sampled value at that time is determined to be an outlier or missing value, and a valid value is obtained only after forward filling, the corresponding mask value for that time is recorded as 1. In other words, a mask value of 1 indicates that the data for that channel at that time is not the original acquired value but a filled value, and a mask value of 0 indicates that the data for that channel at that time is the original acquired valid value.
[0050] The output of the lighting demand prediction sub-model is the normalized value of the lighting power required to maintain the target illuminance at the next time step.
[0051] In this embodiment of the invention, the thermal gain prediction sub-model is set as a long short-term memory network model; The input features of the thermal gain prediction sub-model include normalized values of solar radiation intensity, curtain opening, outdoor temperature, indoor temperature, lighting power, and an input feature validity mask. The output of the thermal gain prediction sub-model is the predicted value of the total indoor thermal gain at the next time step.
[0052] In this embodiment of the invention, the training loss function used by the thermal gain prediction sub-model during training includes a physical consistency regularization term, which is used to apply a positive penalty when the indoor thermal gain prediction value decreases as the solar radiation intensity increases.
[0053] In this embodiment of the invention, the comprehensive energy consumption is the sum of the lighting power demand and the air conditioning power. The air conditioning power is calculated based on the indoor heat gain prediction and the air conditioning energy efficiency ratio. The air conditioning energy efficiency ratio is determined based on the measured outdoor temperature and the unit's partial load rate.
[0054] For example, the core idea of a Long Short-Term Memory (LSTM) network unit is to control the flow of information through three gating structures. The forget gate determines how much information from the previous cell state is retained in the current state; the input gate determines how much of the current input information is written into the cell state; and the output gate determines how much information from the current cell state is output to the hidden state. Through the synergistic effect of these three gating structures, the LSM network can selectively remember long-term dependent information and forget irrelevant information.
[0055] In one specific implementation, the Long Short-Term Memory (LSTM) network unit operates as follows. Let the input feature vector at the current time step be x. τ The hidden state in the previous time step was h. τ-1 The cell state at the previous time step was C. τ-1 First, the forget gate calculates a value between 0 and 1 based on the current input and the previous hidden state, controlling the degree to which the previous cell state is forgotten. Second, the input gate calculates a value between 0 and 1 based on the current input and the previous hidden state, controlling the degree to which new information is written, and simultaneously generates candidate cell states using a hyperbolic tangent function. Then, the current cell state is obtained by adding the previous cell state (after being filtered by the forget gate) to the candidate cell states (after being filtered by the input gate). Finally, the output gate calculates a value between 0 and 1 based on the current input and the previous hidden state, controlling which information from the cell state is output to the current hidden state.
[0056] Through the above calculations, the hidden state at the current time step contains information from both long-term memory and the current input, serving as the output of the short-term memory network unit at this time step. After multiple time steps of recursion, the hidden state at the final step captures the temporal features of the entire input sequence, serving as the input to subsequent fully connected layers.
[0057] Through the gating mechanism of the aforementioned long short-term memory network units, this step enables the network to extract temporal correlations from historical sequences.
[0058] The overall data flow of the dual long short-term memory network sub-model cascade architecture of the present invention is as follows.
[0059] The first level is the lighting demand prediction sub-model. For each candidate strategy, an input tensor containing information about the candidate curtain opening is input into the lighting demand prediction sub-model, and the model outputs the predicted lighting power required to achieve the target illuminance under that curtain opening.
[0060] The second level is the thermal gain prediction sub-model. This model receives the predicted lighting power output from the first level as additional input features, combines it with an input tensor containing candidate curtain opening information, and outputs the predicted total indoor thermal gain under this strategy.
[0061] A data transfer relationship is established between the two-level models: the output of the lighting demand prediction sub-model plays two roles simultaneously—first, it is directly included in the overall energy consumption as a component of lighting energy consumption; second, after normalization, it serves as the input feature of the heat gain prediction sub-model to calculate indoor heat gain, which is then converted into air conditioning power through the energy efficiency ratio model. This forms a complete causal link from lighting power to heat gain, then to air conditioning power, and finally to overall energy consumption.
[0062] During the inference phase of each control cycle, forward propagation is performed in the following order: First, construct the input tensor of the lighting demand prediction sub-model for each candidate strategy, where the curtain opening feature is replaced with the candidate value; Second, perform forward propagation of the lighting demand prediction sub-model to obtain the lighting power prediction value; Third, construct the input tensor of the thermal gain prediction sub-model, where the curtain opening feature is replaced with the same candidate value, and the lighting power prediction value is used as an additional input; Fourth, perform forward propagation of the thermal gain prediction sub-model to obtain the thermal gain prediction value.
[0063] Through the above dual-model cascade architecture design, this step realizes synchronous closed-loop coupling calculation between lighting power and thermal gain, which is different from the hierarchical optimization strategy of existing technologies.
[0064] The input feature vector of the lighting demand prediction sub-model at the current moment includes the following seven features: normalized average outdoor illuminance, normalized average solar radiation intensity, normalized average curtain opening, normalized average indoor illuminance, cyclic time-encoded sine component, cyclic time-encoded cosine component, and input feature validity mask. Among these, the cyclic time-encoded feature maps the number of hours in a day to a continuous numerical range using sine and cosine functions, incorporating prior knowledge of circadian rhythms and avoiding numerical discontinuities at time boundaries.
[0065] For each time step, the input tensor of the lighting demand prediction sub-model is the sequence of the above 7-dimensional feature vectors from the past 30 time steps.
[0066] The output target of the lighting demand prediction sub-model is defined as the normalized value of the lighting power required to maintain a target illuminance of 500 lux at the next time step. This value is taken directly from the actual measured lighting power in historical data during the training phase.
[0067] The network structure of the lighting demand prediction sub-model is as follows: the input layer receives a 30x7 dimensional tensor; the first long short-term memory network has 64 units and outputs the hidden state at each time step; the first dropout layer has a dropout rate of 0.2; the second long short-term memory network has 32 units and outputs only the hidden state at the last step; the second dropout layer has a dropout rate of 0.2; the fully connected layer has 16 neurons and uses a modified linear unit activation function; the output layer has 1 neuron, uses a linear activation function, and directly outputs the normalized prediction value.
[0068] This sub-model uses mean squared error as the training loss function, which is the arithmetic mean of the squares of the differences between the predicted and true values. The optimizer uses an adaptive moment estimator with a learning rate of 0.001.
[0069] In a specific example, the components of the input feature vector for a training sample are 0.65, 0.52, 0.80, 0.42, 0.383, -0.924, and 0, respectively. After forward propagation of the model, the normalized predicted output value is 0.285. After inverse normalization, the predicted lighting power demand is 2422.5 watts, approximately equal to 2.42 kilowatts.
[0070] By introducing long short-term memory network temporal modeling and cyclic time encoding, this sub-model can predict the direction of changes in lighting demand in advance by utilizing the light change trend over the past 30 minutes.
[0071] The input feature vector of the thermal gain prediction sub-model at the current time includes the following six features: normalized average solar radiation intensity, normalized average curtain opening, normalized outdoor temperature, normalized indoor temperature, normalized actual lighting power, and an input feature validity mask. Among these, the normalized actual lighting power uses historical measured values during the training phase and the lighting power prediction value output by the lighting demand prediction sub-model during the inference phase.
[0072] For each time step, the input tensor of the thermal gain prediction sub-model is the above 6-dimensional feature vector sequence of the past 30 time steps.
[0073] The output target of the thermal gain prediction sub-model is defined as the true value of the total indoor thermal gain at the next time step, and the regression is trained directly on the original dimensions.
[0074] The network structure of the thermal gain prediction sub-model is as follows: the input layer receives a 30x6 dimensional tensor; the first layer of long short-term memory network has 64 units and outputs the hidden state at each time step; the second layer of long short-term memory network has 64 units and only outputs the hidden state at the last step; the fully connected layer has 32 neurons and uses a modified linear unit activation function; the output layer has 1 neuron, uses a linear activation function, and directly outputs the thermal gain prediction value.
[0075] To avoid pure data-driven models producing predictions that violate thermodynamic laws under extreme conditions, this sub-model introduces a physical consistency regularization term into the loss function. This regularization term is based on fundamental thermodynamic laws: under the same conditions, an increase in solar radiation intensity cannot lead to a decrease in indoor thermal gain. Specifically, it is implemented by automatically differentiating the model to calculate the partial derivative of the predicted output with respect to the input radiation characteristics. If the partial derivative is negative, a positive penalty is applied; if the partial derivative is non-negative, no penalty is applied. This penalty term, weighted by the regularization term's weighting coefficients, is added to the mean squared error loss term to form the total loss function.
[0076] This physical consistency regularization term guides the network weights towards a physically reasonable parameter space for updates, enhancing the extrapolation reliability of the model under rare weather conditions. The regularization term weight coefficients are determined through cross-validation.
[0077] During the inference phase of each control cycle, the data transfer of the dual-model cascade is performed according to the following steps.
[0078] The first step is forward propagation of the lighting demand prediction sub-model. For each candidate strategy, the input tensor of the lighting demand prediction sub-model under that strategy is constructed, where the curtain opening features at all times are uniformly replaced with the curtain opening values of that candidate strategy. The forward propagation of the lighting demand prediction sub-model is performed, and the normalized lighting power prediction value is output. Then, it is inversely normalized to obtain the actual lighting power value.
[0079] The second step involves the dual use of lighting power. This actual lighting power value serves two purposes: first, it is directly included in the overall energy consumption as a component of lighting energy consumption; second, after normalization, it is used as an input feature of the thermal gain prediction sub-model.
[0080] The third step is forward propagation of the thermal gain prediction sub-model. Construct the input tensor of the thermal gain prediction sub-model under this strategy, where the curtain opening characteristics at all times are uniformly replaced with the curtain opening values of this candidate strategy, and the lighting power characteristics at all times are uniformly replaced with the normalized lighting power prediction values mentioned above. Perform forward propagation of the thermal gain prediction sub-model and output the predicted total indoor thermal gain value.
[0081] The mathematical expression of the above data transmission relationship is as follows: the predicted lighting power value and the predicted thermal gain value are calculated sequentially by the above two-level models in the same forward propagation batch, rather than in a serial mode where the former is calculated first and the latter is calculated after the optimization decision is completed.
[0082] This step clearly defines the data transfer timing and closed-loop coupling relationship between the two models, establishing the essential difference between this application and the existing layered optimization strategies at the architectural level.
[0083] This embodiment collected complete operational data of the target building for 90 consecutive days, covering summer and autumn, including sunny, cloudy, overcast, and rainy weather. After preprocessing, it formed 129,600 observation values at different times.
[0084] Since Long Short-Term Memory (LSTM) networks require continuous sequences, the dataset is divided into the following chronological order: the first 60 days of data are used as the training set for weight updates; the middle 15 days of data are used as the validation set for hyperparameter tuning and early stopping decisions; and the last 15 days of data are used as the test set for final generalization performance evaluation.
[0085] This embodiment performs a grid search on key hyperparameters. The searched hyperparameters include: input sequence length (range 10 to 40); number of hidden layer units (range 32 to 256); number of Long Short-Term Memory (LSTM) layers (range 1 to 3); dropout rate (range 0.0 to 0.5); learning rate (range 0.0001 to 0.005); and batch size (range 16 to 128). For each set of hyperparameters, training is performed on the training set for up to 200 iterations. The mean squared error is calculated on the validation set every 10 iterations, and the group with the lowest mean squared error on the validation set is selected as the optimal configuration.
[0086] The calibration results are as follows: the optimal value for the input sequence length is 30; the optimal value for the number of hidden layer units is 64; the optimal value for the number of long short-term memory network layers is 2; the optimal value for the dropout rate of the lighting demand prediction sub-model is 0.2; the optimal value for the dropout rate of the thermal gain prediction sub-model is 0.0; the optimal value for the learning rate is 0.001; and the optimal value for the batch size is 32.
[0087] All weight matrices are initialized uniformly using Xavier, and all bias vectors are initialized to all zeros. Training is terminated using early stopping: if the validation set loss does not decrease within 15 consecutive iterations and the relative improvement is less than 0.0001, training is terminated and the weight snapshot at the point where the validation set loss is lowest is restored.
[0088] On the test set, the normalized root mean square error of the lighting demand prediction sub-model was 0.031, corresponding to a power error of approximately 263.5 watts after inverse normalization, with a relative error of approximately 3.1%. The root mean square error of the thermal gain prediction sub-model was 412 watts, with a relative error of approximately 8.2%. Both accuracy values meet engineering control requirements.
[0089] This step, through rigorous offline data partitioning, grid search hyperparameter calibration, Xavier initialization, and early stopping, ensures that the two long short-term memory network sub-models possess good generalization performance and numerical stability. After training, the model parameters are frozen and encapsulated as a prediction inference engine, which can be invoked in real-time every 5-minute control cycle for subsequent steps.
[0090] S103. Construct the input features of the thermal gain prediction sub-model based on the predicted lighting power demand value, and construct the input of the thermal gain prediction sub-model together with environmental data and equipment operation data to obtain the indoor thermal gain prediction value. Based on the obtained lighting power demand forecast, this value is divided by the maximum lighting power of 8500 watts to obtain the normalized lighting power value. Multi-dimensional historical sequence data, including normalized solar radiation intensity, normalized curtain opening, normalized outdoor temperature, normalized indoor temperature, and input feature validity mask, are extracted from environmental data and equipment operation data. The above-mentioned lighting power normalized value and multi-dimensional historical sequence data are aligned on the same time axis to form an input tensor containing 6 feature dimensions. This input tensor is input into the pre-trained thermal gain prediction sub-model. After forward propagation, the model outputs the indoor thermal gain prediction value for the next time step, which is measured in watts. In a specific example, when the candidate strategy is a curtain opening of 60% and a target illuminance of 450 lux, the predicted lighting power demand is 2416.2 watts, which is normalized to 0.284. Together with the normalized values of solar radiation (0.48), outdoor temperature (0.52), and indoor temperature (0.45), the input tensor is constructed, and the output thermal gain of the thermal gain prediction sub-model is 6120 watts.
[0091] It is worth noting that the normalized values of lighting power, indoor illuminance, outdoor illuminance, solar radiation intensity, indoor temperature, curtain opening, lighting power, and air conditioning load can be used together as inputs to construct the thermal gain prediction sub-model. Alternatively, one or more types of data from environmental data and equipment operation data can be used together with the normalized values of lighting power as inputs to construct the thermal gain prediction sub-model.
[0092] S104. Calculate the comprehensive energy consumption based on the predicted value of lighting power demand and the predicted value of indoor heat gain. Under the conditions of meeting the preset illuminance constraints, preset temperature constraints and preset glare constraints, determine the optimal control strategy with the goal of minimizing comprehensive energy consumption. The optimal control strategy includes the curtain opening degree, target illuminance and temperature setpoint.
[0093] In this embodiment of the invention, the preset glare constraint is evaluated by a glare risk index proxy model; the glare risk index is calculated based on the equivalent light transmittance of the roller blind, the measured value of solar radiation intensity, the measured value of outdoor illuminance, and the value of lighting power demand; when all candidate control strategies do not meet the preset glare constraint, a multi-level constraint relaxation mechanism is executed in sequence.
[0094] In this embodiment of the invention, determining the optimal control strategy with the goal of minimizing overall energy consumption includes: Multiple candidate control strategies are generated, each candidate control strategy including a combination of candidate values for curtain opening, candidate values for target illuminance, and candidate values for temperature setpoint; Calculate the overall energy consumption for each candidate control strategy; Among the candidate control strategies that meet the preset illuminance constraints, preset temperature constraints, and preset glare constraints, the candidate control strategy with the minimum overall energy consumption is selected as the optimal control strategy.
[0095] In this embodiment of the invention, determining the optimal control strategy with the goal of minimizing overall energy consumption further includes: A penalty term for changes in control action is introduced into the objective function. The penalty term for changes in control action includes the square penalty for changes in curtain opening, the square penalty for changes in temperature setpoint, and the square penalty for changes in target illuminance. The optimal control strategy is the candidate control strategy that minimizes the sum of the total energy consumption and the penalty term for changes in control action.
[0096] For example, in this embodiment, the three control variables are defined and discretized as follows.
[0097] The first control variable is the curtain opening degree, which takes the value as an integer percentage between 0% and 100%. In this embodiment, the discretization step size is set to 10%, generating a candidate value set of 11 values: 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100%. The step size is determined based on the fact that the positioning accuracy of the electric roller blind motor is typically ±3%, and a step size of 10% can effectively distinguish different shading effects without generating an overly dense strategy that would lead to excessive computational burden.
[0098] The second control variable is the target illuminance setpoint. In this embodiment, the target illuminance setpoint is discretized into three levels: 350 lux corresponds to the energy-saving mode, 450 lux corresponds to the normal mode, and 550 lux corresponds to the high visual comfort mode.
[0099] The third control variable is the air conditioning temperature setpoint. In this embodiment, it is discretized into four levels: 23 degrees Celsius, 24 degrees Celsius, 25 degrees Celsius, and 26 degrees Celsius. In this embodiment, the air conditioning system is in cooling mode; the lower the temperature setpoint, the higher the energy consumption. 24 degrees Celsius is the default comfortable temperature in this embodiment, and 25 degrees Celsius is the recommended energy-saving value.
[0100] Through the discretization process described above, the originally continuous space of optimization variables is compressed into a finite set. The total number of candidate strategies is equal to 132, which is calculated by multiplying the number of candidate curtain opening (11), the number of candidate target illuminance (3), and the number of candidate temperature settings (4).
[0101] In this embodiment, a Cartesian product operation is performed on three discrete sets to obtain a set of all candidate control strategies within the current control cycle. The p-th candidate strategy is defined as a triple containing the curtain opening value, the target illuminance value, and the temperature setpoint. Here, p is the candidate strategy index, ranging from 1 to 132.
[0102] In a specific example, the first candidate strategy is 0% curtain opening, a target illuminance of 350 lux, and a temperature setting of 25 degrees Celsius; the last candidate strategy is 100% curtain opening, a target illuminance of 550 lux, and a temperature setting of 23 degrees Celsius. All 132 strategies are stored in the candidate strategy list in the order of curtain opening priority, target illuminance second, and temperature setting last.
[0103] On current mainstream edge computing hardware, forward inference prediction and energy consumption assessment are performed on 132 strategies one by one, with a total time of less than 200 milliseconds, meeting the real-time requirement of a 5-minute control cycle.
[0104] For the p-th candidate strategy, the calculation of lighting power requirements is divided into the following steps.
[0105] The first step is to construct the input tensor for the lighting demand prediction sub-model. Environmental features from the past 30 time periods are extracted from the time-series database. All curtain opening features from the input sequence are then replaced with the normalized form of the current candidate curtain opening values. The normalization method involves dividing the candidate curtain opening percentage value by 100.
[0106] The second step is to perform forward propagation of the lighting demand prediction sub-model. The constructed input tensor is input into the lighting demand prediction sub-model, and the model outputs a normalized lighting power prediction value under a target of 500 lux. This normalized prediction value is multiplied by the difference between the maximum and minimum lighting power values, and then added to the minimum lighting power value to obtain the actual power value. In this embodiment, the maximum lighting power is 8500 watts, and the minimum is 0 watts.
[0107] The third step is target illuminance scaling correction. Since the lighting demand prediction sub-model is trained under a fixed target illuminance of 500 lux, a linear scaling method for luminous flux is used to correct cases where the candidate target illuminance is not equal to 500 lux. Specifically, the predicted lighting power at 500 lux is multiplied by the ratio of the candidate target illuminance to 500 lux, and then multiplied by the overall efficiency correction coefficient of the luminaire dimming driver. This correction coefficient is calibrated through offline bench experiments: 0.98 when the target illuminance is greater than or equal to 500 lux, 0.96 when the target illuminance is between 400 and 500 lux, and 0.95 when the target illuminance is less than 400 lux.
[0108] The fourth step is to construct the input tensor for the thermal gain prediction sub-model. The lighting power demand value is divided by the maximum lighting power of 8500 watts to obtain a normalized value. Environmental features from the past 30 time periods are extracted from the time-series database. The curtain opening features at all times in the input sequence are replaced with the normalized values of the candidate curtain openings, and the lighting power features at all times are replaced with the normalized values of the aforementioned lighting power demand.
[0109] The fifth step is to perform forward propagation of the thermal gain prediction sub-model. The constructed thermal gain prediction sub-model is input into the tensor, and the model outputs the predicted total indoor thermal gain for the next time step, in watts.
[0110] In a specific example, the candidate strategy selected is a curtain opening of 60%, a target illuminance of 450 lux, and a temperature setpoint of 24 degrees Celsius. The normalized power prediction value output by the lighting demand prediction sub-model at 500 lux is 0.329, which becomes 2796.5 watts after inverse normalization. The efficiency correction coefficient corresponding to the target illuminance of 450 lux is 0.96, which, after scaling calculation, yields a lighting power demand of 2416.2 watts. The normalized value of curtain opening (0.6) and lighting power (0.284), along with the current measured normalized values of solar radiation (0.48), outdoor temperature (0.52), and indoor temperature (0.45), are used to construct an input tensor. This tensor is then input into the thermal gain prediction sub-model, which outputs a thermal gain of 6120 watts.
[0111] This embodiment uses a bivariate quadratic polynomial model based on outdoor temperature and partial load rate to dynamically estimate the air conditioning energy efficiency ratio under the current operating conditions.
[0112] The mathematical expression of the model is: the energy efficiency ratio equals a constant term, plus the outdoor temperature multiplied by the first-order coefficient of the outdoor temperature, plus the part load factor multiplied by the first-order coefficient of the part load factor, plus the square of the outdoor temperature multiplied by the second-order coefficient of the outdoor temperature, plus the square of the part load factor multiplied by the second-order coefficient of the part load factor, plus the product of the outdoor temperature and the part load factor multiplied by the cross term coefficient.
[0113] In the above model, the coefficients are as follows: constant term coefficient is -2.15, outdoor temperature linear term coefficient is 0.28, part load factor linear term coefficient is 4.20, outdoor temperature quadratic term coefficient is -0.0035, part load factor quadratic term coefficient is -2.10, and cross term coefficient is -0.045.
[0114] The partial load factor is calculated by dividing the predicted indoor heat gain by the nominal cooling capacity of the air conditioning unit. If the ratio is greater than 1, it is taken as 1. In this embodiment, the nominal cooling capacity of the air conditioning unit is 150,000 watts.
[0115] After obtaining the energy efficiency ratio, the air conditioning power consumption under the p-th candidate strategy is calculated using the following formula: Air conditioning power equals the predicted indoor heat gain divided by the energy efficiency ratio. The total energy consumption is the sum of lighting power demand and air conditioning power.
[0116] Assuming the current measured outdoor temperature is 33 degrees Celsius, the calculated heat gain is 6120 watts. The partial load factor is 6120 divided by 150000, which equals 0.0408. Substituting these values into the energy efficiency ratio (EER) model, the EER is approximately 3.39. The air conditioning power consumption is 6120 divided by 3.39, which is approximately 1805.3 watts. The total energy consumption is the sum of lighting power (2416.2 watts) and air conditioning power (1805.3 watts), equaling 4221.5 watts.
[0117] According to the standard values and allowable deviation range of illuminance for office buildings in GB 50034-2022 "Standard for Lighting Design of Buildings", this embodiment sets the average illuminance on the work surface to be within the range of 350 lux to 550 lux. The target illuminance values for each candidate strategy in this embodiment are 350, 450, and 550 lux, all of which meet this constraint.
[0118] According to the acceptable temperature range for summer cooling conditions in ASHRAE Standard 55-2020 Thermal Comfort Standard, this embodiment sets the indoor air temperature to be between 23 and 26 degrees Celsius. The temperature setpoints for each candidate strategy in this embodiment are 23, 24, 25, and 26 degrees Celsius, all of which meet this constraint.
[0119] The glare risk index defined in this embodiment is calculated as follows. First, the equivalent light transmittance of the roller blind is calculated. This value is equal to the light transmittance when the roller blind is fully closed plus one minus the light transmittance when the roller blind is fully closed multiplied by the square of the normalized value of the blind opening. In this embodiment, a blackout electric roller blind is used, and the light transmittance when the roller blind is fully closed is calibrated to 0.02.
[0120] The formula for calculating the glare risk index is: constant term 0.15, plus equivalent transmittance multiplied by the current solar radiation intensity multiplied by a coefficient of 0.0012, plus equivalent transmittance multiplied by the current outdoor illuminance divided by 1000 multiplied by a coefficient of 0.25, plus lighting power demand divided by the maximum lighting power of 8500 multiplied by a coefficient of 0.15.
[0121] The meanings of the above coefficients are as follows: constant term coefficient 0.15, solar radiation term coefficient 0.0012 square meters per watt, outdoor illuminance term coefficient 0.25 per kilolux, and lighting power term coefficient 0.15.
[0122] The calibration method for each coefficient is as follows: The calibration data comes from on-site measurement experiments conducted on a typical sunny day at this building. On the experimental day, in addition to the conventional sensors, a luminance imager was additionally set up at a height of 1.2 meters above the ground and 3 meters from the window, recording a high dynamic range luminance distribution image every 5 minutes. Afterwards, the standard solar glare probability value at each moment was calculated using software. Simultaneously, the solar radiation intensity, outdoor illuminance, curtain opening, and lighting power data from the conventional sensors were recorded. A total of 120 valid data points were collected. With the standard solar glare probability value as the dependent variable, a multiple linear regression was performed with three independent variables, and the coefficient of determination was 0.89.
[0123] The glare constraint is: the glare risk index is less than or equal to 0.38.
[0124] For each candidate strategy, three constraints are checked sequentially: illuminance constraint, temperature constraint, and glare constraint. In this embodiment, the illuminance and temperature constraints are automatically satisfied by the discretization design, and the actual selection is mainly based on the glare constraint.
[0125] If the glare risk index of all candidate strategies exceeds the hard threshold, the constraint relaxation mechanism is triggered. In this embodiment, the priority order is set as follows: temperature constraint takes precedence over illuminance constraint, and illuminance constraint takes precedence over glare constraint.
[0126] The constraint relaxation mechanism is implemented in the following steps: Level 1 relaxation relaxes the glare threshold from 0.38 to 0.45; Level 2 relaxation further relaxes the threshold to 0.50 if no feasible strategy is found after Level 1 relaxation; Level 3 relaxation completely removes the glare constraint if no feasible strategy is found after Level 2 relaxation, retaining only the illuminance and temperature constraints.
[0127] This mechanism ensures that the system can still output feasible strategies under extreme conditions, prioritizing temperature constraints and then illuminance constraints.
[0128] The optimal strategy for this control cycle is determined by minimizing the following objective function. The objective function consists of five parts: lighting power demand, air conditioning power, the square penalty for changes in curtain opening, the square penalty for changes in temperature setting, and the square penalty for changes in target illuminance.
[0129] The weighting coefficients for lighting power demand and air conditioning power demand are both 1.0. The penalty coefficient for changes in curtain opening is 5.0 watts per square percent, the penalty coefficient for changes in temperature is 80 watts per square degree Celsius, and the penalty coefficient for changes in illuminance is 0.01 watts per square lux.
[0130] The aforementioned penalty coefficients were determined using a questionnaire survey. Questionnaires were distributed to 20 office workers to collect their comfort ratings for sudden changes in curtain position, temperature (sudden drop or rise), and illuminance. After calculating the weighting of each abrupt change factor, the penalty coefficients were derived.
[0131] In a specific example, suppose the control variables executed in the previous control cycle were a curtain opening of 50%, a temperature setpoint of 24.5 degrees Celsius, and a target illuminance of 500 lux. The current candidate strategy is a curtain opening of 60%, a target illuminance of 450 lux, and a temperature setpoint of 24 degrees Celsius. The lighting power is 2416.2 watts, the air conditioning power is 1805.3 watts, and the total energy consumption is 4221.5 watts. The curtain penalty is 5.0 multiplied by the square of 60 minus 50 (in parentheses), equaling 500.0 watts. The temperature penalty is 80 multiplied by the square of 24 minus 24.5 (in parentheses), equaling 20.0 watts. The illuminance penalty is 0.01 multiplied by the square of 450 minus 500 (in parentheses), equaling 25.0 watts. The total objective function value is 4766.5 watts.
[0132] This embodiment employs a two-step look-ahead rolling optimization scheme. The first-step cost function value is calculated based on the prediction at the current time. The second-step cost function value is obtained by recursively calculating the cost function value by substituting the strategy from the first step into the cascaded prediction model. The cumulative cost is the first-step cost function value plus the second-step cost function value multiplied by the future discount factor. In this embodiment, the future discount factor is set to 0.8.
[0133] The computational complexity of two-step lookahead is 132 times 132, which equals 17,424 evaluations. It takes about 1.5 seconds to complete on an edge device, meeting the time window of a 5-minute control cycle.
[0134] After rolling optimization calculations, the optimal strategy is the candidate strategy that minimizes the cumulative cost, and this candidate strategy must satisfy the constraints. The optimal strategy is a triple, which includes the optimal curtain opening, the optimal target illuminance, and the optimal temperature setpoint.
[0135] This step ultimately outputs three engineering parameters for the optimal strategy: curtain opening degree (in percentage); target illuminance (in lux); and temperature setpoint (in degrees Celsius).
[0136] The three engineering parameters of the optimal strategy are converted into control signals for the corresponding actuators.
[0137] First, the curtain opening percentage is sent to the electric roller blind motor driver, with the signal type being "target position." The conversion method is as follows: divide the curtain opening percentage by 100, then multiply by the full travel time of the electric roller blind motor from fully closed to fully open. In this embodiment, this full travel time has been measured and calibrated to be 45 seconds. This instruction is sent via the Modbus RTU communication protocol, with function code 0x06, register address 0x0001, and the written value being the integer value obtained by multiplying the curtain opening percentage by 10.
[0138] Second, the target illuminance is sent to the LED dimming driver as an illuminance setpoint signal. This command is sent via a digitally addressable lighting interface communication protocol, carrying the target illuminance value. The LED dimming driver has a built-in proportional-integral-derivative control loop that automatically adjusts the pulse width modulation duty cycle to approach the target value based on the currently measured illuminance.
[0139] Third, the temperature setpoint is sent to the air conditioning terminal controller, with the signal type being the set temperature value. This instruction is sent via the BACnet MS / TP communication protocol, written to the attribute service, with the object type being AnalogOutput, and directly writing the temperature setpoint in floating-point format.
[0140] In a specific example, if the optimal strategy is a curtain opening of 60%, a target illuminance of 450 lux, and a temperature setpoint of 24 degrees Celsius, then the curtain instruction is a travel time of 27 seconds, written with the value 600 via the Modbus protocol; the lighting instruction is sent with an illuminance setpoint of 450 via the Digital Addressable Lighting Interface Protocol; and the air conditioning instruction is written with 24.0 via the BACnet protocol.
[0141] In this embodiment, the synchronous execution time window is set to 100 milliseconds. Specifically, after completing the optimization calculation, the edge smart gateway prepares three control commands in memory but does not immediately issue them; it waits until 100 milliseconds before the next control cycle arrives, then simultaneously issues the commands through its respective communication channels; each actuator begins its action within 100 milliseconds after receiving the command, achieving synchronous execution. In this embodiment, the control cycle is 300 seconds.
[0142] By using a synchronous execution window design, the three instructions take effect at almost the same time, avoiding intermediate transition states caused by timing overlap. It's worth noting that the three parameters can also be converted into the execution parameters corresponding to the controller of the relevant device, depending on the actual situation.
[0143] Optionally, after executing the optimal control strategy, the actual operating data is continuously collected to calculate the lighting power prediction deviation between the predicted lighting power demand and the actual lighting power value, as well as the heat gain prediction deviation between the predicted indoor heat gain and the actual indoor heat gain value. When the absolute value of the lighting power prediction deviation exceeds the first threshold or the absolute value of the heat gain prediction deviation exceeds the second threshold, the incremental update of the solar thermal load prediction model is triggered.
[0144] The system can continuously record the following actual operating data in real time at a sampling period of 60 seconds: average illuminance of the actual working surface (lux); average indoor temperature (degrees Celsius); actual lighting circuit power (watts); actual air conditioning power (watts); and actual curtain opening degree (percentage). All actual data are filtered and outlier removed, and then aligned with historical predicted values in chronological order.
[0145] The relative deviation of lighting power prediction is defined as: the actual lighting power value minus the predicted lighting power value, divided by the predicted lighting power value, and then multiplied by 100%, in percentage form. The predicted lighting power value is the prediction of the lighting demand prediction sub-model for the current time from the previous sampling time.
[0146] The relative deviation of the thermal gain prediction is defined as: the actual indoor thermal gain value minus the predicted thermal gain value, divided by the predicted thermal gain value, and then multiplied by 100%, in percentage form. The predicted thermal gain value is the prediction of the thermal gain prediction sub-model for the current time from the previous sampling time.
[0147] The method for inversely calculating the actual indoor heat gain is as follows: Obtain the actual measured chilled water volumetric flow rate and the actual supply and return water temperature difference at the current moment, and substitute them into the air conditioning heat load calculation formula. This formula is: the specific heat capacity of water at constant pressure multiplied by the density of water, then multiplied by the chilled water volumetric flow rate and the supply and return water temperature difference. This method directly inversely calculates the heat gain based on actual physical measurements and does not rely on iterative calculations of the energy efficiency ratio model.
[0148] In one specific example, the predicted lighting power at a certain moment was 2420 watts, while the actual measured value was 2560 watts, resulting in a relative deviation of approximately 5.79% in the lighting power prediction. The predicted thermal gain was 6120 watts, while the actual calculated value was 5880 watts, resulting in a relative deviation of approximately -3.92% in the thermal gain prediction.
[0149] This embodiment sets up a dual triggering mechanism.
[0150] Trigger condition 1 is short-term cumulative deviation triggering. If the absolute value of the relative deviation of the lighting power prediction is greater than 10% or the absolute value of the relative deviation of the thermal gain prediction is greater than 15% within 12 consecutive sampling times, and this continues for more than 10 consecutive sampling times, an online fine-tuning incremental update will be triggered immediately.
[0151] Trigger condition 2 is a periodic retraining trigger. Every Monday at 02:00 AM, an offline retraining is automatically triggered, using all accumulated historical data up to the current moment, including at least the data from the most recent 30 days, to completely retrain both sub-models.
[0152] When trigger condition 1 is met, the system performs an online incremental update. The specific process is as follows: extract the most recent 500 sample points from the time series database as the incremental training set, and perform preprocessing and normalization according to step S101; load the weights of the currently deployed long short-term memory network model as the initial weights, set the incremental learning rate to 0.0001, the batch size to 16, and the maximum number of fine-tuning iterations to 10; after fine-tuning, verify whether the prediction bias has fallen below the threshold on the most recent 200 samples that have not participated in training. If the verification is successful, replace the currently deployed model with the updated model weights.
[0153] The scheduled retraining performed every Monday morning follows the exact same process as pretraining: extracting historical data from the last 90 days, re-performing dataset partitioning, grid search hyperparameter calibration, full training, and early stopping validation. After retraining, the model is replaced before 5:00 AM that same day, without affecting normal control operations.
[0154] This step enables the predictive model to adaptively track changes in building usage patterns, equipment performance degradation, and seasonal climate pattern changes through a dual-trigger update mechanism.
[0155] like Figure 2 As shown, the present invention provides a shading and lighting linkage control system based on solar thermal load prediction, comprising: The data acquisition module 201 is used to collect environmental data and equipment operation data through a sensor network deployed indoors and outdoors. The environmental data and equipment operation data include indoor illuminance, outdoor illuminance, solar radiation intensity, indoor temperature, curtain opening, lighting power and air conditioning load data. The lighting prediction module 202 is used to build and pre-train the solar and thermal load prediction model. The solar and thermal load prediction model includes a cascaded lighting demand prediction sub-model and a thermal gain prediction sub-model. Environmental data and equipment operation data are input into the lighting demand prediction sub-model to obtain the predicted value of lighting power demand. The thermal gain prediction module 203 is used to construct the input features of the thermal gain prediction sub-model based on the predicted lighting power demand value. It is used together with environmental data and equipment operation data to construct the input of the thermal gain prediction sub-model and obtain the indoor thermal gain prediction value. The linkage control module 204 is used to calculate the comprehensive energy consumption based on the predicted value of lighting power demand and the predicted value of indoor heat gain. Under the conditions of meeting the preset illuminance constraints, preset temperature constraints and preset glare constraints, the optimal control strategy is determined with the goal of minimizing the comprehensive energy consumption. The optimal control strategy includes the curtain opening degree, target illuminance and temperature setpoint.
[0156] The shading and lighting linkage control system based on solar thermal load prediction in this application embodiment can execute the method in any embodiment, and the principle and effect are the same.
[0157] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0158] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0159] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0160] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0161] Although preferred embodiments of the present 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 the embodiments of the present application.
[0162] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.
[0163] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for coordinated control of shading and lighting based on solar thermal load prediction, characterized in that, include: Environmental data and equipment operation data are collected through an indoor and outdoor sensor network, including indoor illuminance, outdoor illuminance, solar radiation intensity, indoor temperature, curtain opening, lighting power, and air conditioning load data. Construct and pre-train a solar and thermal load prediction model; the solar and thermal load prediction model includes a cascaded lighting demand prediction sub-model and a thermal gain prediction sub-model. The environmental data and equipment operation data are input into the lighting demand prediction sub-model to obtain the predicted value of lighting power demand. The input features of the thermal gain prediction sub-model are constructed based on the predicted lighting power demand value. Together with the environmental data and equipment operation data, the input of the thermal gain prediction sub-model is constructed to obtain the indoor thermal gain prediction value. The comprehensive energy consumption is calculated based on the predicted lighting power demand and the predicted indoor thermal gain. Under the conditions of satisfying the preset illuminance constraints, preset temperature constraints and preset glare constraints, the optimal control strategy is determined with the goal of minimizing the comprehensive energy consumption. The optimal control strategy includes the curtain opening degree, target illuminance and temperature setpoint.
2. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, Before inputting the environmental data and equipment operation data into the lighting demand prediction sub-model, the following steps are also included: The environmental data and equipment operation data are preprocessed, including sliding window filtering, outlier detection and removal based on the 3σ criterion, forward filling and masking of missing values, and / or minimum-maximum normalization.
3. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, The lighting demand prediction sub-model is set as a long short-term memory network model; The input features of the lighting demand prediction sub-model include normalized outdoor illuminance, normalized solar radiation intensity, normalized curtain opening, normalized indoor illuminance, cyclic time-coded sine component, cyclic time-coded cosine component, and input feature validity mask. The output of the lighting demand prediction sub-model is the normalized value of the lighting power required to maintain the target illuminance at the next time step.
4. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, The thermal gain prediction sub-model is set as a long short-term memory network model; The input features of the thermal gain prediction sub-model include normalized values of solar radiation intensity, normalized values of curtain opening, normalized values of outdoor temperature, normalized values of indoor temperature, normalized values of lighting power, and a mask for the validity of input features. The output of the thermal gain prediction sub-model is the predicted value of the total indoor thermal gain at the next time step.
5. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, The training loss function used in the thermal gain prediction sub-model during training includes a physical consistency regularization term, which is used to apply a positive penalty when the indoor thermal gain prediction value decreases as the solar radiation intensity increases.
6. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, The comprehensive energy consumption is the sum of the lighting power demand and the air conditioning power demand. The air conditioning power demand is calculated based on the indoor heat gain prediction and the air conditioning energy efficiency ratio. The air conditioning energy efficiency ratio is determined based on the measured outdoor temperature and the unit's partial load rate.
7. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, The preset glare constraint is evaluated using a glare risk index proxy model. The glare risk index is calculated based on the equivalent light transmittance of the roller blind, the measured value of solar radiation intensity, the measured value of outdoor illuminance, and the value of lighting power demand. When all candidate control strategies fail to meet the preset glare constraint, a multi-level constraint relaxation mechanism is executed sequentially.
8. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, Determining the optimal control strategy with the goal of minimizing the overall energy consumption includes: Multiple candidate control strategies are generated, each candidate control strategy including a combination of candidate values for curtain opening, candidate values for target illuminance, and candidate values for temperature setpoint; Calculate the overall energy consumption for each candidate control strategy; Among the candidate control strategies that satisfy the preset illuminance constraint, the preset temperature constraint, and the preset glare constraint, the candidate control strategy with the minimum overall energy consumption is selected as the optimal control strategy.
9. The shading and lighting linkage control method based on solar thermal load prediction according to claim 1, characterized in that, Determining the optimal control strategy with the goal of minimizing the overall energy consumption also includes: A penalty term for changes in control action is introduced into the objective function; the penalty term for changes in control action includes the square penalty for changes in curtain opening, the square penalty for changes in temperature setpoint, and the square penalty for changes in target illuminance; the optimal control strategy is the candidate control strategy that minimizes the sum of the overall energy consumption and the penalty term for changes in control action.
10. A shading and lighting linkage control system based on solar thermal load prediction, characterized in that, include: The data acquisition module is used to collect environmental data and equipment operation data through a sensor network deployed indoors and outdoors. The environmental data and equipment operation data include indoor illuminance, outdoor illuminance, solar radiation intensity, indoor temperature, curtain opening degree, lighting power and air conditioning load data. The lighting prediction module is used to build and pre-train a solar and thermal load prediction model. The solar and thermal load prediction model includes a cascaded lighting demand prediction sub-model and a thermal gain prediction sub-model. The environmental data and equipment operation data are input into the lighting demand prediction sub-model to obtain the predicted value of lighting power demand. The thermal gain prediction module is used to construct the input features of the thermal gain prediction sub-model based on the predicted lighting power demand value, and to construct the input of the thermal gain prediction sub-model together with the environmental data and equipment operation data to obtain the indoor thermal gain prediction value. The linkage control module is used to calculate the comprehensive energy consumption based on the predicted lighting power demand and the predicted indoor heat gain. Under the conditions of satisfying the preset illuminance constraints, preset temperature constraints and preset glare constraints, the optimal control strategy is determined with the goal of minimizing the comprehensive energy consumption. The optimal control strategy includes the curtain opening degree, target illuminance and temperature setpoint.
