Multi-objective optimization energy-saving control method for fan heater and fan heater control board

Abstract

This invention provides a multi-objective optimized energy-saving control method and control board for a heater. By transmitting a narrowband sweep frequency excitation signal and simultaneously acquiring reflected waves at the moment of near-field communication wake-up, a time-series channel fingerprint vector reflecting environmental parameters such as device location and medium humidity is generated. Combined with a high-precision system timestamp, a dynamic session token valid within a specific time window is calculated using a hash algorithm. This token, along with periodically broadcast beacon frames, enables remote device authentication. A session key is generated based on the authentication result using a key derivation function, thereby establishing a secure encrypted communication link under multi-device concurrent access. This invention effectively eliminates static environmental interference, improves the uniqueness, tamper resistance, and timeliness of authentication, and adapts to the identity authentication and secure communication needs in complex near-field environments.

Description

Technical Field

[0001] This invention relates to the field of energy-saving control technology for heating, ventilation and air conditioning systems, and particularly to a multi-objective optimization energy-saving control method for heaters and a heater control board. Background Technology

[0002] Currently, in the field of energy saving and comfort control of HVAC systems, especially in the area of ​​warm air blowers, most mainstream technical solutions adopt a multi-objective optimization approach based on preset static models. Traditional energy-saving strategy generation modules generally rely on PMV thermal comfort indices, ASHRAE standard comfort models, typical temperature control curves, or offline-trained big data neural network models to calculate control parameters. These methods typically pre-set the user comfort range and energy-saving target weights, and provide the optimal control strategy based on static or semi-static models during equipment operation. The industry trend is to promote a refined synergy between "minimizing energy consumption" and "maximizing user thermal comfort" by introducing more types of environmental and physiological sensor data, improving the feature dimensions of models, and introducing machine learning or data mining techniques. However, existing technologies have the following prominent problems in practical applications: (1) Most systems emphasize building thermal comfort models offline and then using them for online control. Their parameters, weights, boundaries, etc. need to be set in advance and cannot be corrected in real time. Individual users have significant differences in their sensitivity to heat and cold, usage habits and pace of life. Fixed models are difficult to accurately reflect actual needs, especially in scenarios where individual experiences are significant, such as at home, where the strategy is significantly lagging behind. (2) Even some intelligent systems with learning capabilities often only use explicit adjustments on the temperature control interface as feedback signals, failing to fully explore the implicit behavioral characteristics of user operations. For example, users' habits of frequently fine-tuning the temperature, preferring to run at a low fan speed, or actively turning off the device at a certain time actually imply individualized thermal regulation patterns, which have not yet been systematically utilized in existing systems; (3) Most current energy-saving optimization directions are aimed at structured environmental variables (outdoor temperature, equipment status), lacking dynamic integration with personal thermal comfort "transfer" capabilities. This makes it difficult to achieve real-time correction and adaptive transfer of strategies in situations with large environmental changes and many interfering factors (such as opening and closing doors in the home, flow of people entering and leaving, etc.). (4) With the large-scale application of the Internet of Things and embedded control boards in heaters, the limitation of computing resources is becoming increasingly prominent. Methods such as using high-dimensional physiological signal sensors, complex neural network inference, or cloud big data training are difficult to operate economically and efficiently at resource-constrained edge devices. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a multi-objective optimization energy-saving control method for a warm air blower and a control board for the warm air blower.

[0004] The technical solution of this invention is implemented as follows: a multi-objective optimization energy-saving control method for a warm air blower, comprising: S1: Acquire four types of key operation events triggered by the user during the operation of the heater and synchronized environmental status data. The four types of key operation events include manually increasing the set temperature, manually decreasing the set temperature, actively shutting down and restarting the device, long-pressing the fan speed setting, and timeout on the temperature control interface. The synchronized environmental status data includes the ambient temperature at the corresponding time, the current power of the device, the running time, and the slope of the heater surface temperature change, in order to generate the original behavior log sequence. S2: Based on the original behavior log sequence, construct a timestamp-weighted behavior response sequence for each type of key operation event, extract the device power decrease rate, the initial rate of ambient temperature drop, and the time interval between the user's subsequent increase operation within a preset time period after manually lowering the set temperature as joint features, and generate a thermal regulation inertia vector with temporal stability through sliding window aggregation processing. S3: For multiple thermal regulation inertia vectors under different ambient temperature ranges, the local density peak clustering algorithm is used to identify the typical comfort response patterns of individual users in each ambient temperature range, and generate a discrete comfort response distribution set containing multiple local cluster centers. S4: Based on the discrete comfort response distribution set, perform incremental Gaussian mixture modeling on the historical local cluster centers, update the mean vector representing the current comfort temperature center and the covariance matrix reflecting the user's tolerance elasticity to temperature fluctuations, so as to generate a dynamic thermal preference fingerprint distribution that evolves over time. S5: The dynamic thermal preference fingerprint distribution is injected into the sparse variational Gaussian process surrogate model as prior knowledge. The joint response surface of the energy consumption reduction per unit time and the fingerprint matching score is fitted using no more than a preset number of induction points to generate an energy-saving-comfort trade-off prediction model. S6: Based on the prediction variance plot of the energy-saving-comfort trade-off prediction model, calculate the next control action parameter that maximizes the decrease in posterior entropy. The control action parameter includes the set temperature fine-tuning amplitude and the duration of maintenance, so as to generate the active sampling control command with the greatest information gain. S7: Execute the active sampling control command to drive the heater to adjust its operating state, and capture in real time whether the user generates new operation feedback and corresponding environmental disturbance data in the new operating state, so as to generate a closed-loop verification dataset containing execution results and new feedback information; S8: Use the closed-loop verification dataset to update the induction point position of the sparse variational Gaussian process surrogate model and the statistical parameters of the dynamic hot preference fingerprint distribution, correct the model's cognitive bias of the user's real preferences, and generate an adaptive optimization control strategy for the next control cycle.

[0005] The present invention also provides a heater control board, which uses the above-mentioned multi-objective optimization energy-saving control method for heaters to perform multi-objective optimization energy-saving control of heaters.

[0006] The multi-objective optimization energy-saving control method and control board for a warm air blower provided by this invention have the following beneficial effects: (1) This invention achieves dynamic capture of individual preferences without any additional sensing hardware or standard thermal comfort formulas by embedding a lightweight behavior log module in the control board of the heater and constructing a time-weighted operation response sequence. By jointly modeling the user's natural operation behavior (such as temperature adjustment, power off, and interface dwell) and the resulting device and environmental responses, a "thermal regulation inertia vector" with time-series stability is generated, which significantly improves the system's sensitivity and accuracy in recognizing the user's subjective comfort intentions. This mechanism transforms human intervention into quantifiable feedback signals, which not only avoids the deployment of complex human physiological parameter acquisition devices, but also enables the system to continuously accumulate personalized data under unsupervised and zero-calibration conditions, greatly reducing deployment costs and usage thresholds. (2) This invention introduces a two-layer clustering and incremental Gaussian hybrid modeling mechanism to construct a "thermal preference fingerprint" distribution that evolves over time, enabling the system to dynamically represent the user's thermal comfort center and fluctuation tolerance, breaking through the technical bottleneck of multi-objective adaptive coordination in traditional recommendation or control strategies. The bottom-layer local density peak clustering can accurately identify the typical response patterns of users in different environmental temperature ranges, while the upper-layer incremental GMM realizes cross-time period preference trend tracking. Its mean and covariance reflect the current optimal set point and temperature fluctuation tolerance range, respectively, forming an interpretable individualized comfort model. After the fingerprint is injected as prior knowledge into the Bayesian optimizer driven by the sparse variational Gaussian process, it can efficiently fit the joint response surface of the dual objectives of energy saving and comfort under limited computing resources, and select the fine-tuning action with the largest information gain through an active sampling strategy, realizing a paradigm shift from "passive response operation" to "active exploration of preferences". This invention significantly improves the foresight and personalization level of the control strategy, ensuring that the user experience is maintained while continuously approaching the optimal energy efficiency boundary. (3) The entire closed-loop learning framework of this invention runs entirely on resource-constrained embedded platforms (such as ARM Cortex-M7 level MCUs). All modules adopt lightweight design, eliminating the need for offline training of deep networks, cloud computing, or large-scale historical datasets, and truly achieving localized, real-time, and low-power online adaptive control. Compared to intelligent temperature control schemes that require the pre-collection of a large amount of user data and centralized training, this invention, through induced point compression, sliding window aggregation, and incremental update mechanisms, controls memory usage and computational overhead within the feasible range of embedded systems while ensuring model expressiveness, supporting long-term uninterrupted autonomous learning. At the same time, the system uses user behavior feedback as the sole input source, forming a complete cognitive closed loop of "execution-observation-correction," which not only enhances the transparency and traceability of control logic but also provides a good architectural foundation for subsequent functional expansion (such as multi-device collaboration and scene migration). Attached Figure Description

[0007] Figure 1 This is a flowchart of the multi-objective optimization energy-saving control method for the heater of the present invention; Figure 2 This is a sub-flowchart of the multi-objective optimization energy-saving control method for the heater of the present invention; Figure 3 This is another sub-flowchart of the multi-objective optimization energy-saving control method for the heater of the present invention. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides a multi-objective optimization energy-saving control method for a heater, specifically including: S1: Acquire four types of key operation events triggered by the user during the operation of the heater and synchronized environmental status data. The four types of key operation events include manually increasing the set temperature, manually decreasing the set temperature, actively shutting down and restarting the device, long-pressing the fan speed setting, and timeout on the temperature control interface. The synchronized environmental status data includes the ambient temperature at the corresponding time, the current power of the device, the running time, and the slope of the heater surface temperature change, in order to generate the original behavior log sequence. S2: Based on the original behavior log sequence, construct a timestamp-weighted behavior response sequence for each type of key operation event, extract the device power decrease rate, the initial rate of ambient temperature drop, and the time interval between the user's subsequent increase operation within a preset time period after manually lowering the set temperature as joint features, and generate a thermal regulation inertia vector with temporal stability through sliding window aggregation processing. S3: For multiple thermal regulation inertia vectors under different ambient temperature ranges, the local density peak clustering algorithm is used to identify the typical comfort response patterns of individual users in each ambient temperature range, and generate a discrete comfort response distribution set containing multiple local cluster centers. S4: Based on the discrete comfort response distribution set, perform incremental Gaussian mixture modeling on the historical local cluster centers, update the mean vector representing the current comfort temperature center and the covariance matrix reflecting the user's tolerance elasticity to temperature fluctuations, so as to generate a dynamic thermal preference fingerprint distribution that evolves over time. S5: The dynamic thermal preference fingerprint distribution is injected into the sparse variational Gaussian process surrogate model as prior knowledge. The joint response surface of the energy consumption reduction per unit time and the fingerprint matching score is fitted using no more than a preset number of induction points to generate an energy-saving-comfort trade-off prediction model. S6: Based on the prediction variance plot of the energy-saving-comfort trade-off prediction model, calculate the next control action parameter that maximizes the decrease in posterior entropy. The control action parameter includes the set temperature fine-tuning amplitude and the duration of maintenance, so as to generate the active sampling control command with the greatest information gain. S7: Execute the active sampling control command to drive the heater to adjust its operating state, and capture in real time whether the user generates new operation feedback and corresponding environmental disturbance data in the new operating state, so as to generate a closed-loop verification dataset containing execution results and new feedback information; S8: Use the closed-loop verification dataset to update the induction point position of the sparse variational Gaussian process surrogate model and the statistical parameters of the dynamic hot preference fingerprint distribution, correct the model's cognitive bias of the user's real preferences, and generate an adaptive optimization control strategy for the next control cycle.

[0011] Step S1: Acquire four types of key operation events triggered by the user during the operation of the heater, along with synchronized environmental status data. The four types of key operation events include manually increasing the set temperature, manually decreasing the set temperature, actively shutting down and restarting the device, long-pressing the fan speed setting, and timeout on the temperature control interface. The synchronized environmental status data includes the ambient temperature at the corresponding moment, the device's current power, operating time, and the slope of the heater surface temperature change, to generate an original behavior log sequence. Specifically, this includes: S1.1: Perform interrupt polling and edge detection processing on the input interface signals and touch interface status of the heater control panel to identify and capture four types of key operation events: manually raising the set temperature, manually lowering the set temperature, actively shutting down and restarting the device, long-pressing the fan speed setting, and timeout of the temperature control interface, and generate operation event trigger markers with microsecond-level timestamps. The goal of performing interrupt polling and edge detection processing on the input interface signals and touch interface status of the heater control panel is to convert manually triggered control actions into precise event markers in real time, ensuring accurate timing alignment of subsequent environmental status data collection. The input objects include debounced digital signals from mechanical button circuits, multi-touch coordinates and pressure data from capacitive touch screens, and the temperature control interface dwell time count value in the system UI status register; Based on the mechanical key input signal, the hardware interrupt trigger condition is set to level flipping and stable for more than 2 sampling cycles after digital debounce filtering, triggering the interrupt service routine to read the key code and map it as a set temperature increase or decrease event; For touch screen input signals, perform a fixed-period polling scan, compare the coordinate change rate of the touch event with a static threshold using an edge detection operator to determine whether the long press action of adjusting the fan speed is completed; if the duration of the touch event exceeds the interface dwell threshold, generate a temperature control interface timeout dwell event flag. For the power control logic of the equipment, monitor the state toggling of the power switch input signal on the control panel and combine it with the system operating mode register to determine the occurrence of an event that actively shuts down or restarts the equipment; After all events are detected, the event type code is bound to the microsecond-level timestamp read by the system's high-precision timer to form an operation event trigger marker, and written into the event FIFO queue to ensure the uniqueness and timing traceability of the trigger marker in subsequent environmental data collection; By using the above-mentioned interrupt polling and edge detection processing methods, the original physical interaction signals are transformed into structured event markers with controllable time precision, thereby achieving low latency and high accuracy in capturing user operations. For example, on a heater control board equipped with a combination of capacitive touchscreen and mechanical buttons, the mechanical button scan cycle is configured to 5ms, the debouncing filter window is set to 2 cycles, the touchscreen polling interval is 8ms, the long press judgment threshold is 2.5s, the temperature control interface dwell timeout threshold is 90s, and the microsecond-level timestamp is obtained by the driver layer calling the hardware timer and read directly in the interrupt context. When the temperature increase button is pressed, the level flips stably after debouncing filtering and exists in the sampling buffer for two consecutive cycles. In the second cycle, a hardware interrupt is triggered, the button code is read in the ISR and mapped to the "manually increase set temperature" event, and the timestamp is recorded. In microseconds, this marker is stored in the event FIFO queue. When a long press on the fan speed adjustment operation reaches a set action threshold for 320 consecutive sampling periods during touchscreen coordinate change rate detection, and the touch duration exceeds 2.5 seconds, the edge detection operator is triggered to generate an event marker, and the timestamp is recorded as follows. Microseconds. The temperature control interface timeout event is triggered when the dwell time count in the UI status register reaches the 90-second threshold, and the recorded timestamp is... Microseconds. In different user testing scenarios, all event markers can be generated within a single cycle after the triggered behavior occurs, and are precisely time-aligned with the corresponding environmental state acquisition, significantly improving the response accuracy and reliability of subsequent hot preference modeling; S1.2: Based on the time point of the operation event trigger mark, synchronously read the digital filter output value of the ambient temperature sensor, the instantaneous power sampling value of the power metering module, the cumulative duration count value of the system running timer, and the differential rate of change value of the heater surface temperature sensor to generate a multi-dimensional synchronous environmental state data set that is strictly time-aligned with the operation event trigger mark. Based on the microsecond-level timestamp of the operation event trigger mark, the digital filtering module of the ambient temperature sensor is called to output a stable temperature sampling value. The digital filtering module performs noise reduction processing on the original ADC sample through a multi-order finite impulse response (FIR) filter, so that the output has low latency and high steady-state accuracy. The instantaneous power sampling channel of the power metering module performs a rapid sampling at the trigger moment. This channel calculates the single-point power by multiplying the calibrated voltage and current. The calculation formula is: in For sampling voltage, For sampling current; The system reads the cumulative duration of the running timer and outputs the running time in milliseconds through a hardware counter. This count value serves as an important index for subsequent association with environmental status and device operation stages. The instantaneous temperature value of the heater surface temperature sensor is acquired, and a first-order backward differential operation is performed to obtain the temperature change rate. The calculation formula is: in The surface temperature of the heater. This is the sampling interval time; The above four types of sampled data are strictly aligned according to the trigger timestamp, and linear interpolation or zero-order hold strategy is used to handle the timing inconsistency of each sampled channel, so as to ensure that each parameter is synchronized at the same trigger time. Through synchronization processing, ambient temperature values, power values, runtime, and temperature change rate from different sources are bound into multi-dimensional vector records, forming a multi-dimensional synchronized environmental state data set that corresponds one-to-one with the operation event trigger marker, thereby achieving accurate matching of data in subsequent encapsulation and serialization stages. For example, in a heater with a built-in ARM Cortex-M7 MCU, when the user manually lowers the set temperature, a trigger flag occurs at a microsecond-level timestamp of 1000000us. The ambient temperature sensor outputs 25.4 degrees Celsius after a third-order FIR filter. The power metering module quickly samples the voltage as 220.5 volts and the current as 2.31 amps. This is then calculated using the formula... The calculated instantaneous power is 509.3555 watts, the cumulative duration of the running timer output is 7,200,000 milliseconds, and the two sampled values ​​of the heater surface temperature sensor are 58.3 degrees Celsius (t) and 58.7 degrees Celsius (t) respectively. Sampling interval It is 0.5 seconds, according to the formula Calculate the rate of temperature change 0.8 degrees Celsius / second. These four data points were time-aligned using zero-order hold and synthesized into a vector [25.4, 509.3555, 7200000, [0.8], and stored in the synchronization environment state data set in a one-to-one correspondence with the trigger marker to verify that this alignment method can significantly improve the context integrity of subsequent behavior logs within the millisecond-level response period after triggering; S1.3: The operation event triggering flag and the multidimensional synchronous environment state data set are structured and encapsulated using a circular buffer storage mechanism. Discrete event type codes and continuous physical quantity parameters are bound into a single record unit to generate original behavior log record entries containing complete context information. S1.4: Perform sliding window aggregation and serialization processing on multiple consecutively generated raw behavior log entries in chronological order, remove invalid noise records caused by sensor jitter, and form a raw behavior log sequence arranged linearly on the timeline, which serves as the direct input basis for constructing the timestamp-weighted behavior response sequence.

[0012] Step S2: Based on the original behavior log sequence, a timestamp-weighted behavior response sequence is constructed for each type of key operation event. The device power decrease rate, the initial rate of ambient temperature drop, and the time interval between the user's subsequent temperature increase operation within a preset time period after manually lowering the set temperature are extracted as joint features. These features are then aggregated through a sliding window to generate a thermal regulation inertia vector with temporal stability. Specifically, this includes: S2.1: Perform timeline alignment processing on each type of key operation event in the original behavior log sequence, and extract data segments before and after a preset time length based on the time of the operation event, generating a time-synchronized data subset containing ambient temperature, current power of the equipment, and the slope of the change in heater surface temperature; S2.2: Based on the time-synchronized data subset, the timestamp weighted calculation of each sampling point in the data segment is performed using an exponential decay function, and data points closer to the time of the operation event are given higher weights to generate a time-weighted state sequence that reflects the intensity of the user's immediate response. An index mapping relationship is established for each data sampling point in the time synchronization data subset. The difference between the sampling time and the reference time of the operation event is calculated as the time offset, and a time series parameter matrix is ​​constructed in seconds. Each time offset in the time series parameter matrix is ​​used as the input variable of the exponential decay function. A weight value sequence is generated according to the preset decay coefficient, which is determined by the system experimental calibration to ensure that the weight reaches its peak near the reference time and decreases with time. Using a power-law model, the time-weighted value is calculated using the following formula: in, For weight values, The attenuation coefficient is... This is the absolute value of the time offset; a weighted product operation is performed on the original values ​​of the slope of change of ambient temperature, equipment power, and heater surface temperature using this sequence of weighted values. The result of the weighted product operation is mapped to the weighted state matrix to form a second-order matrix structure with the physical quantity dimension on the horizontal and the time sampling point on the vertical, so as to support the matrix operation efficiency of subsequent feature extraction. Normalization is performed on each column of the weighted state matrix to map data of different dimensions to the [0,1] interval, ensuring that the weight ratio between physical quantities remains consistent in the weighted state sequence, thereby generating a time-weighted state sequence with consistent dimension characteristics. By processing the time-synchronized data subset using an exponential decay function, the time-axis aligned data fragments from the previous step are converted into a continuous time-weighted state sequence, thereby achieving a quantitative expression of the instantaneous response intensity at the moment a user operation event occurs. For example, when the heater is set to a constant temperature of 22°C and operating at 1200W, and the user manually lowers the set temperature by 2°C, the system extracts a 60-second time synchronization data subset before and after the temperature change, with a sampling interval of 1 second, forming 121 sampling points. The time offset sequence consists of integer values ​​from -60 to +60 seconds, and an attenuation coefficient is selected. = Calculate the weight values ​​according to the formula; for example, the weight values ​​at the reference time. = Weight value at an offset of 30 seconds = The ambient temperature (21.8℃), power (1150W), and surface temperature change slope (-0.12℃ / s) at the corresponding time points were multiplied by weight values ​​to obtain weighted temperature (21.8℃), weighted power (256.15), and weighted slope (-0.026). After normalization, these values ​​were mapped to the standard space to obtain numerical inputs of 0.99, 0.22, and 0.45. The time-weighted state sequence after processing all sampling points maximized at the reference time and decreased with a smooth curve. In the verification, this sequence, when used for feature extraction, significantly improved the stability and discriminability of the thermal regulation inertia vector in cluster analysis. S2.3: Perform differential feature extraction operation on the time-weighted state sequence, calculate the equipment power decrease rate and the initial rate of ambient temperature drop within a preset time period for the manually lowered set temperature event, and combine the time interval of the user's subsequent increase operation to generate a multi-dimensional joint feature group characterizing the user's thermal regulation dynamic characteristics. S2.4: The multidimensional joint feature group is processed by time-series aggregation using the sliding window algorithm. During the sliding process, the mean vector and variance matrix of the features within the window are calculated to eliminate random noise interference and generate a local feature aggregation vector with time-series smoothing characteristics. S2.5: Perform a normalization mapping transformation based on the local feature aggregation vector to uniformly map the power change rate, temperature change rate and time interval of different dimensions to the standard feature space, and output a thermal regulation inertia vector that has temporal stability and can be directly used for cluster analysis.

[0013] like Figure 2As shown, step S3 involves using a local density peak clustering algorithm to identify typical comfort response patterns of individual users within different ambient temperature ranges, targeting multiple thermal regulation inertia vectors, and generating a discrete comfort response distribution set containing multiple local cluster centers. Specifically, this includes: S3.1: Based on the ambient temperature component in the thermal regulation inertia vector, perform interval mapping and data splitting processing to divide the continuous ambient temperature values ​​into multiple preset temperature threshold intervals, and aggregate the thermal regulation inertia vectors falling into the same temperature threshold interval into independent environmental context data subsets to generate clustering vector groups isolated by temperature intervals. Based on the ambient temperature components in the normalized thermal regulation inertial vector dataset, the temperature value corresponding to each vector is read as the input for interval mapping, and a deterministic interval mapping rule is established in combination with the temperature threshold partitioning table configured by the system. The continuous temperature values ​​are discretized according to the mapping rules. An interval determination operation is then performed between the input temperature value and the preset lower and upper limits of the interval. The determination formula can be expressed as follows: in, The ambient temperature component is extracted from the thermally regulated inertia vector. and The first The lower and upper limits of a temperature range, Mark the interval matching results; When the determination result is true, the corresponding thermal adjustment inertia vector is written into the buffer unit of that interval to realize data diversion; Index binding is performed on the vectors within each temperature range cache unit, and the split vectors and their range numbers are encapsulated into environmental context data subset objects to ensure that the partitioned data is independent in subsequent clustering operations; A consistency check is performed on all environmental context data subsets to remove vector records with missing temperature components or outliers. Through this process, the global thermal regulation inertia vector set generated in the previous step is transformed into a group of vectors to be clustered, isolated by temperature range, thus realizing the structured preparation of context data with environmental temperature as the key dimension. By using interval mapping and data splitting, the thermal regulation inertia vector result from the previous step is transformed into a clustering vector group for the environmental temperature partition, achieving the expected technical effects of environmental context isolation and improved clustering input accuracy. For example, in the normalized thermal regulation inertia vector set output by the ambient temperature acquisition module of the heater, each vector contains components such as ambient temperature, power change rate, initial temperature drop velocity, and time interval. When acquiring the ambient temperature component, interval determination is performed according to the preset threshold table [16.0~19.9], [20.0~23.9], and [24.0~27.0]℃. For the vector with a temperature value of 18.7℃, the input formula is... When the interval parameter = and = When the judgment condition is met, an interval matching marker is generated. The vector is stored in the cache with interval number 1. For the vector with a temperature value of 22.5℃, the result matches the interval [20.0~23.9], and an interval marker is generated. The data is stored in cache number 2. After splitting all vectors, three independent environmental context data subsets are formed, each containing 5-8 vector records. In subsequent clustering, vectors within each subset only calculate distance and density with other vectors in the same interval, avoiding cross-interval temperature interference, and significantly improving the speed of cluster center identification. S3.2: For each thermally regulated inertial vector in the group of vectors to be clustered, calculate its Euclidean distance matrix with all other thermally regulated inertial vectors, and convert the Euclidean distance matrix into a local density index based on the Gaussian kernel function to generate a sequence of local density values ​​characterizing the density of samples around each thermally regulated inertial vector. Based on the thermally regulated inertia vectors contained in the group of vectors to be clustered, which are isolated by temperature range, all standardized feature components of the vector elements are obtained by traversing their index positions one by one, and used as the input variable set for distance calculation. For each target thermally regulated inertial vector, the Euclidean distance calculation module is called to accumulate the corresponding feature components of the vector with all other thermally regulated inertial vectors in the same interval according to the element difference sum of squares structure, and apply the square root operation to the sum of squares to generate a complete Euclidean distance matrix, in which each element of the matrix corresponds to a set of distance values ​​between samples. A Gaussian kernel function transformation is performed on the Euclidean distance matrix. During the calculation, the distance values ​​are used as input to the exponential function and mapped in the following form: These are the element values ​​of the Euclidean distance matrix. The kernel function bandwidth parameter is determined through empirical formulas or by searching for extrema on the validation set. Perform row vector summation on the transformed Gaussian kernel matrix, and use the cumulative result of all kernel values ​​in the row corresponding to each target thermally regulated inertial vector as the local density index of the vector, and generate a sequence of local density values ​​in the vector group in sequence. Through the above Gaussian kernel function mapping and density accumulation processing, the standardized feature spatial distance information of the previous step is transformed into a local density value sequence that characterizes the density of the samples, thus realizing the preliminary identification basis of the comfort response pattern based on spatial distribution. For example, in a group of vectors to be clustered within an ambient temperature range of 20–23℃, each thermal regulation inertia vector is a three-dimensional feature vector [power change rate, temperature change rate, time interval]. Let a target vector be [0.15, 0.08, 120]. Calculate the Euclidean distances to the other five vectors in the same group. For example, the distance to vector [0.10, 0.05, 100] is... After obtaining the numerical value, input it into the Gaussian kernel function, assuming the bandwidth parameter... The calculated kernel value is approximately 0.92. After performing kernel transformation on all five distances, the kernel values ​​are summed to obtain a local density value of approximately 4.63, which serves as an indicator of the sample density of this vector within that temperature range. Verification shows that inputting this local density value into the subsequent calculation process for cutoff distance and decision score can significantly improve the accuracy of comfort response mode center point selection. S3.3: Based on the local density value sequence, calculate the minimum distance from each thermally regulated inertial vector to its neighbor with a higher local density value as a cutoff distance index, and multiply the local density value with the cutoff distance index to generate a decision score, so as to select a set of candidate cluster center points whose decision scores exceed a preset threshold. Based on the input of the local density value sequence, a neighborhood search and comparison operation is performed on each thermally regulated inertial vector. All neighbors with local density values ​​higher than the current vector are selected in the index space of the Euclidean distance matrix and their distance set is calculated. The minimum distance in the set is extracted as the cutoff distance index. The cutoff distance index is productted with its local density value to generate a decision score sequence that distinguishes between extreme and non-extreme points. The decision scores are calculated using the following formula: in, For decision scores, This represents the local density value of the current thermally regulated inertia vector. This is the minimum Euclidean distance from the vector to its neighbors with higher local density values; The decision score sequence is compared with a preset threshold, and all thermally adjusted inertial vectors with decision scores greater than the threshold are selected. Their indices and spatial coordinates are then encapsulated into a set of candidate cluster center points. By using the chain-like processing method described above, the local density value sequence and the truncation distance information are fused into a cluster center discrimination index, thereby realizing the automatic extraction of candidate center points; For example, in a group of vectors to be clustered within an ambient temperature range of 20–23℃, the local density value of a certain vector is... In its Euclidean distance matrix, the minimum distance to neighbors with a local density value higher than 8.6 is... The cutoff distance metric is set to... The decision score D is calculated using the formula: The calculation result is When the preset threshold is set to When the decision score of the vector exceeds the threshold, it is marked as a candidate cluster center. This calculation and comparison are performed sequentially on all thermally regulated inertia vectors within the interval, and finally five candidate cluster centers are output. Their spatial distribution covers the main dense data areas within the interval. Subsequent clustering and assignment verification shows that these centers can significantly improve the stability and accuracy of comfort response pattern recognition. S3.4: Initialize the clustering assignment process using the candidate cluster center point set, and assign each remaining thermally adjusted inertial vector to the cluster category of the nearest neighbor with a higher local density value, until all thermally adjusted inertial vectors have been assigned a category, so as to generate a preliminary clustering result set containing complete category labels; S3.5: For each cluster in the preliminary clustering result set, extract the geometric centroid coordinates of all thermal regulation inertia vectors in that cluster as local cluster centers, and encapsulate the local cluster centers corresponding to all temperature threshold intervals into structured data objects to generate a discrete comfort response distribution set containing multiple local cluster centers.

[0014] like Figure 3 As shown, step S4 involves: based on the discrete comfort response distribution set, performing incremental Gaussian mixture modeling on the historical local cluster centers to update the mean vector representing the current comfort temperature center and the covariance matrix reflecting the user's tolerance elasticity to temperature fluctuations, thereby generating a dynamic thermal preference fingerprint distribution that evolves over time. Specifically, this includes: S4.1: Obtain data on multiple local cluster centers in the discrete comfort response distribution set, and perform weighted preprocessing on the historical local cluster centers based on the time decay weight function to generate a weighted historical cluster center sequence containing time-related weights; Using local cluster center data in the discrete comfort response distribution set as input, each local cluster center data is appended with its generation timestamp information, and arranged in ascending order of timestamps to form a historical cluster center time sequence chain; Based on this time-series chain, a time decay weight function is constructed, which assigns higher weights to cluster centers that are closer to the current time and gradually reduces the weights of historical cluster centers with longer time intervals. The time decay weight function is exponential, and the weight value of each cluster center is calculated using the following formula: in, For the first The weight values ​​of each cluster center For the current moment and the first The time difference between the generation times of each cluster center The time decay constant; The above weight calculation is performed on the historical cluster center data, and the obtained weights are multiplied element-wise with the coordinate components of each dimension of the cluster center to achieve time-effective weighted processing of discrete data. The weighted historical cluster center data is encapsulated into structured record units containing coordinate values, weight values, and timestamps, and output as a weighted historical cluster center sequence, providing basic data for timeliness constraints in subsequent Gaussian mixture modeling; By applying the time decay weighting function, the discrete comfort response distribution set is transformed into a weighted historical cluster center sequence that reflects the evolution trajectory of user thermal preferences. This enables a quantitative characterization of the timeliness differences of historical data and lays a weighted foundation for modeling the dynamic thermal preference fingerprint distribution. For example, during the operation of the warm air blower system, a local cluster center is generated every 10 minutes, with the temperature component ranging from 19.5℃ to 24.1℃ and the power change rate ranging from [missing value]. 0.3 to Between 0.8 and 0.8. Assume a time decay constant. The time difference is 120 minutes. If the time of generation of a cluster center is only 20 minutes from the current time, then the time difference is... Calculate the weight value for a time interval of 20 minutes: ≈ For cluster centers located 100 minutes from the current time, the weight value is... ≈ The aforementioned weights are then multiplied by the temperature component and power change rate corresponding to the cluster center, for example, the temperature component is 22.0℃ and the power change rate is... With a weight of 0.5 and a 20-minute weighting, the weighted temperature component is approximately 22.0 × 0.846 ≈ 18.612, and the power change rate is... 0.5 × 0.846 ≈ 0.423; Under a 100-minute weighting, the weighted temperature component is approximately 22.0 × 0.435 ≈ 9.57, and the power change rate is... 0.5 × 0.435 ≈ 0.217. After weighting all historical cluster centers, a weighted historical cluster center sequence containing temperature components, power change rate, weight values, and timestamps is formed. This provides Gaussian mixture model with data input that combines spatial features and time decay constraints. This sequence significantly improves the sensitivity of the dynamic thermal preference fingerprint distribution to recent changes in user comfort boundaries in subsequent modeling. S4.2: Based on the weighted historical cluster center sequence, the Gaussian mixture model parameters are iteratively calculated using the expectation-maximization algorithm to generate an initial set of Gaussian components that characterize the probability density of typical user comfort modes under different ambient temperature ranges. Based on the multidimensional coordinate data and corresponding time-sensitive weights in the weighted historical cluster center sequence, a data matrix is ​​constructed as the input to the Gaussian mixture model to ensure that the samples in different temperature ranges are reasonably proportioned in subsequent calculations according to the decay weights. The Gaussian mixture model is initialized for the data matrix. The number of components is determined based on the number of temperature ranges and the density of cluster centers within each range. The mean vector is initialized to the geometric centroid of the cluster centers in each range, and the covariance matrix is ​​initialized to the overall covariance of the sample vectors within the range. The initialized model is input into the iterative framework of the Expectation-Maximization algorithm. In the E-step, the posterior probability of each sample corresponding to different Gaussian components is calculated using the formula: in, Let be the posterior probability of the i-th weighted historical cluster center belonging to the k-th Gaussian component. The weight of the k-th Gaussian component. Let be a Gaussian probability density function. Let i be the weighted historical cluster center vector. Let be the mean vector of the k-th Gaussian component. Let be the covariance matrix of the k-th Gaussian component. For the traversal index of all Gaussian components; In the M-step, the mean vector, covariance matrix, and component weights of each Gaussian component are updated using the posterior probability. The update formula for the mean vector is: The formula for updating the covariance matrix is: By maximizing the expectation iteratively until the log-likelihood value converges or the preset number of iterations is reached, the weighted historical cluster center samples under different ambient temperature ranges are mapped to a set of initial Gaussian components, whose probability density accurately describes the distribution characteristics of the user's typical comfort mode in each range. Through iterative calculation using the expectation-maximization algorithm, the weighted historical cluster center sequence from the previous step is transformed into an initial Gaussian component set that meets the requirements of statistical stability, thereby achieving accurate modeling of the probability density of comfort modes within different ambient temperature ranges. For example, during a 30-day operation cycle of the heater system, 15, 18, and 12 local cluster centers were collected for each ambient temperature range (e.g., 16–19℃, 20–23℃, 24–27℃), respectively, with time decay weights set to 0.85, 0.78, and 0.91. A three-dimensional input data matrix was constructed, with each row containing the mean of comfort response features, the rate of power change, and the rate of temperature change. The number of components in the Gaussian mixture model was set to be the same as the number of intervals, i.e., 3 components. The initial mean vector was the geometric centroid of each interval, and the initial covariance matrix was calculated from the overall covariance of the samples within the interval. In the expectation-maximization iteration, the E-step posterior probability calculation relied on the above formula. Assuming a sample x=(0.3,0.5,0.2) in the first interval, the mean... =(0.25,0.52,0.21), covariance The matrix is ​​a diagonal matrix (0.01, 0.015, 0.012). The Gaussian density values ​​are calculated according to the set kernel function and updated in conjunction with the component weights. Value. In the M-step, using The mean vector is updated to (0.26, 0.51, 0.205). After updating the covariance matrix, the variance of the temperature change rate dimension decreases, and the stability of the power change rate improves. After 20 iterations, the log-likelihood value stabilizes at -120.4. The initial Gaussian component set output includes the mean, covariance, and weight of each component. The application effect significantly improves the model's ability to fit the evolution of the user comfort boundary in subsequent recursive Bayesian updates. S4.3: Based on the initial Gaussian component set, a recursive Bayesian update mechanism is used to perform online correction processing on the mean vector and covariance matrix of the Gaussian mixture model to generate an updated Gaussian component parameter set that reflects the user's latest thermal regulation behavior. Based on the mean vector and covariance matrix parameters in the initial Gaussian component set, the prior input structure of the recursive Bayesian update mechanism is set, and the Gaussian mixture model generated in the previous control cycle is used as the prediction benchmark for the current time. In each update cycle, the multidimensional Mahalanobis distance of the newly added local cluster center samples on each Gaussian component is calculated, and the association probability weight sequence of the sample to each component is generated. Based on the aforementioned correlation probability weight sequence, a weighted correction is performed on the mean vector of each Gaussian component, using a recursive update formula of the following form: in, This is the updated mean vector at the current time. This is the mean vector of the previous time step. For the new observation sample vector, Here are the Bayesian gain coefficients. After the mean is updated, an element-wise recursive smoothing strategy is used on the covariance matrix. The outer product is calculated based on the difference between the new sample and the updated mean vector, and then corrected using the following formula: In the above formula, The updated covariance matrix has an outer product term used to capture the variation in the dispersion of the latest samples relative to the mean. The corrected mean vector and covariance matrix are written back to the corresponding Gaussian components to form the updated Gaussian component parameter set. The weight of each component is normalized by the product of the prior weight and the likelihood value of the new sample to ensure the stability of the weight distribution. By using a recursive Bayesian update mechanism, the initial Gaussian components from the previous step are transformed into an updated set of Gaussian component parameters that reflect the user's latest thermal regulation behavior, thereby achieving real-time correction of the comfort temperature center and temperature fluctuation tolerance elasticity. For example, in a scenario where a heater control board is deployed, the initial Gaussian component set contains three components with mean vectors of [20.5, 19.8], [22.0, 20.1], and [24.3, 21.5], and covariance matrices of [[0.4,0.1],[0.1, 0.3]], [[0.5,0.12],[0.12,0.45]], and [[0.6,0.14],[0.14,0.5]]. The new observation sample vector is [21.2, 20.0]. After calculating the Mahalanobis distance of the sample on each component, the association probability weights are obtained as [0.6,0.3, 0.1]. For the first component, the mean vector is updated to [20.92, 19.88] and the covariance matrix is ​​updated to [[0.42, 0.11], [0.11, 0.32]] according to the above recursive formula. After the update, the mean and covariance of the three components are corrected respectively, and the weights are kept to a sum of 1 after likelihood normalization. This process is executed every 5 minutes on the control board MCU. After 10 consecutive updates, the trajectory of the mean vector gradually approaches the latest comfort temperature center shown by the user in actual operation. The trace value of the covariance matrix shows a significant decrease in the stage of stable user habits, and the model's predicted comfort boundary and dynamic energy consumption matching ability are significantly improved. S4.4: Based on the updated Gaussian component parameter set, extract the mean vector of each Gaussian component as the current comfort temperature center index and calculate the trace of the covariance matrix as the temperature fluctuation tolerance elasticity index to generate structured dynamic thermal preference fingerprint distribution data.

[0015] Step S5: The dynamic thermal preference fingerprint distribution is injected as prior knowledge into a sparse variational Gaussian process surrogate model. A joint response surface of the energy consumption decrease per unit time and the fingerprint matching score is fitted using no more than a preset number of induced points to generate an energy-saving-comfort tradeoff prediction model. Specifically, this includes: S5.1: Based on the mean vector and covariance matrix in the dynamic thermal preference fingerprint distribution, the Latin hypercube sampling algorithm is used to generate an initial induction point candidate set in a low-dimensional operating space, and the initial induction point candidate set is mapped to the prior kernel function parameters that characterize the user's comfort temperature center and temperature fluctuation tolerance elasticity, so as to complete the initial configuration of the sparse variational Gaussian process proxy model. Based on the mean vector and covariance matrix in the dynamic thermal preference fingerprint distribution, two core parameters representing the user's comfort temperature center and temperature fluctuation tolerance elasticity are extracted by analyzing the fingerprint distribution structure and used as prior inputs. Within the low-dimensional operating space, the Latin hypercube sampling algorithm is used to perform uniform hierarchical sampling on the two control dimensions of set temperature and duration of maintenance, generating an initial induction point candidate set covering the entire operating space. The coordinate components of each induction point in the candidate set are parameter-mapped, the mean vector of the comfort temperature center is mapped to the kernel function center parameter, and the trace of the covariance matrix corresponding to the temperature fluctuation tolerance elasticity is mapped to the kernel function scaling coefficient. Normalization is performed during the mapping process to unify the temperature center and elastic parameters of different dimensions to a standard scale, so as to ensure the numerical stability of the kernel function in multi-task modeling. The induced point set after mapping is embedded into the initialization stage of the sparse variational Gaussian process surrogate model as the basis point set for constructing the prior kernel function, ensuring that the model has an initial cognition that matches the current user's real preferences when fitting the energy-saving and comfort goals. Through the above processing method, the dynamic thermal preference fingerprint distribution of the previous step is transformed into the prior kernel function parameters that can directly drive the initialization of the sparse variational Gaussian process surrogate model, so as to achieve rapid convergence and high adaptability of the model in the prediction of energy saving-comfort trade-off. For example, in the operating scenario of an embedded heater control board, the mean vector of the dynamic thermal preference fingerprint distribution is [22.6℃, 630s], and the trace value of the covariance matrix is ​​1.4 (℃²-second²). The low-dimensional operating space is set with a temperature range of 21.0℃ to 24.0℃, a step size of 0.2℃, and a duration range of 300s to 900s, with a step size of 30s. The Latin hypercube sampling algorithm is used to generate an initial candidate set of 20 induction points. During the sampling process, a normalization transformation is performed on the set temperature components, mapping 22.6℃ to the kernel function center parameter of 0.8, and a linear scaling mapping is performed on the covariance trace of 1.4 to the kernel function scaling factor of 0.35. For Latin hypercube sampling, a uniformly covered sequence 21.0, 21.8, 22.6, 23.4, 24.0 is generated in the temperature dimension, and a uniformly covered sequence is generated in the duration dimension. And perform prior kernel parameter mapping for each sampling combination point. The kernel function uses a squared exponential kernel, and its covariance function is of the form: ,in The scaling coefficients obtained from the mapping. The length scale is derived from the comfort temperature center. Current control action parameters, Another control parameter. After this step is completed, the kernel function structure and induced point distribution during the initialization of the surrogate model are consistent with the user's latest thermal preferences. In practice, the number of convergence iterations in subsequent gradient optimization is significantly reduced, and the deviation between energy saving and comfort prediction is greatly reduced. S5.2: The prior kernel function parameters are associated with the historical operation data in the original behavior log sequence. The Cramer divergence between the likelihood function of the observed data and the variational distribution is calculated using a variational inference algorithm to generate a variational lower bound objective function value for measuring the degree of approximation of the current induced point position to the true posterior distribution. S5.3: Based on the variational lower bound objective function value, execute the gradient ascent optimization strategy to iteratively update the position coordinates and number of the initial candidate set of induced points, eliminate redundant induced points and adjust the distribution density of key induced points on the joint response surface, so as to generate the optimal set of induced points that meets the preset quantity constraints. S5.4: Reconstruct the covariance matrix structure of the sparse variational Gaussian process surrogate model using the optimal set of induction points, and perform multi-task joint modeling with the energy consumption reduction per unit time as the first output dimension and the fingerprint matching score as the second output dimension to generate a joint response surface containing the mean prediction vector and the variance uncertainty matrix. Based on the determined optimal set of induction points, the spatial coordinates of each induction point in the set are matrix-combined with the prior kernel function parameters in the dynamic thermal preference fingerprint distribution to form a joint parameter input matrix; The covariance function structure is reconstructed for the input matrix, mapping the correlation between each induced point to the matrix block structure required for multi-task modeling, corresponding to the two output dimensions of energy saving target and comfort target respectively; The energy consumption reduction per unit time in the first output dimension is modeled using a squared exponential kernel, and the fingerprint matching score in the second output dimension is fitted using a Matern kernel function. A cross-correlation term is introduced into the covariance matrix to characterize the coupling relationship between the two targets. Perform inverse decomposition operation under sparse variational inference on the above multi-core covariance matrix to obtain the mean prediction vector and variance uncertainty matrix of each output dimension at the induction point position; The mean prediction vector and variance uncertainty matrix are encapsulated into a joint response surface data structure, and the original control action parameter index is associated with each surface node for subsequent global scanning process. By using a multi-task joint modeling process, the optimal set of induction points from the previous step is transformed into a two-dimensional surface that can simultaneously characterize the energy-saving efficiency and thermal comfort matching response, thus realizing the core construction of the energy-saving-comfort trade-off prediction model. For example, in the operating data of a heater, the optimal induction point set contains 15 two-dimensional operational space coordinate points. The first dimension represents the set temperature adjustment range, and the second dimension represents the duration of the adjustment. The mean vector of the dynamic heat preference fingerprint distribution is... The covariance matrix is The covariance matrix is ​​embedded into the dual-kernel structure of the sparse variational Gaussian process surrogate model, and the length scale of the squared exponential kernel parameter is set to... For energy-saving purposes, the length scale of the Matern kernel parameter is set to... For comfort objectives. When constructing the multi-task covariance matrix, cross-core coefficients are introduced. This represents the linear coupling strength between the two targets. After performing sparse variational inference, the mean predicted value of the first output dimension at the first induced point is... The variance uncertainty is The second output dimension's mean prediction value at the same position is... Uncertainty is Similar bi-objective response predictions were obtained at all nodes on the global surface. It was verified that the surface can significantly improve the balance between energy saving and comfort in the control strategy of the heater, and the prediction results remain stable under multiple ambient temperature scenarios. S5.5: Perform full-domain scanning and feature extraction on the joint response surface, analyze the weighted coupling relationship between the predicted energy efficiency and the predicted comfort satisfaction under different control action parameters, so as to generate an energy-saving and comfort trade-off prediction model with online evolution capability; Based on the mean prediction vector and variance uncertainty matrix in the joint response surface, a full-domain scanning data framework covering the control parameter space is constructed to ensure that the predicted values ​​of energy efficiency and comfort satisfaction are comparable within the same parameter domain. For each combination of control action parameters obtained by scanning, energy efficiency feature analysis is performed on its first output dimension to extract the local gradient information of energy consumption decrease per unit time, which is used as a sensitivity index of energy saving target. Dynamic feature extraction of comfort satisfaction prediction values ​​is performed on the second output dimension with the same combination of control action parameters. The change of fingerprint matching score under neighborhood perturbation is calculated as a stability index of comfort target. A weighted coupling function is used to numerically fuse the energy-saving target sensitivity index and the comfort target stability index. The weighting coefficients are determined based on the ratio of the trace value of the covariance matrix of the dynamic thermal preference fingerprint distribution to the variance of energy consumption prediction. The fusion calculation formula is as follows: in, The weighted coupling value, These are the weighting coefficients. As a sensitive value for energy-saving targets, The target stable value is for comfort. Global feature extraction is performed on the fused weighted coupling values ​​to identify the control action parameter region with the largest positive gradient and prediction variance below a preset threshold, which serves as the efficient execution domain of the energy-saving-comfort trade-off model. Through feature extraction and fusion operations, the joint response surface output from the previous step is transformed into an energy-saving and comfort trade-off prediction model with online evolution capabilities, thereby achieving a dynamic balance between energy saving and comfort in the control strategy. For example, within the range of values ​​for fine-tuning the set temperature of the heater [ Within a range of [0.5℃, 0.5℃] and a duration of [1 min, 5 min], the predicted energy efficiency and comfort satisfaction values ​​for each parameter combination were obtained through a full-domain scan. For a specific combination (temperature adjustment +0.3℃, duration 2 min), the energy efficiency target sensitivity was 0.12 kWh / h, and the comfort target stability was 0.85. The trace value of the covariance matrix of the dynamic thermal preference fingerprint distribution was 0.04, and the energy consumption prediction variance was 0.02. Weighting coefficients were calculated based on these values. = = Substituting into the fusion formula: = The weighted coupling value of this combination lies within the efficient execution domain of the scan space. The model predicts that this combination can significantly improve energy efficiency and maintain high comfort under the current environment, with the prediction variance within a safe range. Validation data after actual implementation of this combination shows a significant reduction in power consumption, and no user feedback regarding cooling was generated before the end of the maintenance period, verifying the robustness and online evolution capability of the prediction model.

[0016] Step S6: Based on the prediction variance plot of the energy-saving-comfort tradeoff prediction model, calculate the next control action parameter that maximizes the decrease in posterior entropy. The control action parameter includes the set temperature fine-tuning amplitude and the duration of maintenance, to generate the active sampling control command with the highest information gain. Specifically, this includes: S6.1: The joint response surface output by the sparse variational Gaussian process surrogate model is discretized by global gridding, and a multidimensional control action space composed of the combination of candidate set temperature fine-tuning amplitude and candidate maintenance duration is extracted. The posterior prediction variance value corresponding to each discrete action point is read to generate a prediction variance map dataset containing the mapping relationship between action parameters and uncertainty measure. S6.2: Based on the predicted variance map dataset, the posterior predicted variance value of each candidate action point is nonlinearly transformed using the Shannon entropy calculation formula to map the high variance region to a high information entropy value, quantify the model knowledge update potential expected after each candidate action is executed, and generate a sequence of expected posterior entropy decrease values ​​that characterize the information gain potential of each action. Based on each candidate action point and its corresponding posterior predicted variance value in the predicted variance map dataset, the input data is processed by vectorization, and an index mapping structure is established to correspond one-to-one between the action parameter coordinates and the predicted variance value to support the subsequent entropy value calculation. The Shannon entropy calculation formula is introduced into the predicted variance value set for nonlinear transformation processing. Specifically, each predicted variance value is converted into a probability quantification factor and then mapped to an information entropy value to reflect the knowledge gain potential of the action after execution. For each candidate action point, a probability normalization operation is performed to ensure that the sum of the probability quantization factors of all action points is one. This process assigns higher probability weights to points with high variance to highlight the importance of their uncertainty. The Shannon entropy formula is used to quantize the normalized probability factor. Calculate the entropy value: in The normalized probability quantization factor. This represents the information entropy value of the corresponding action; The expected posterior entropy decrease is calculated by combining the above entropy sequence with the model prediction data. By comparing the coupling degree between the expected decrease in the model prediction variance after each action and the entropy value, a sequence of expected posterior entropy decrease values ​​representing the potential of information gain is formed. The expected value sequence of posterior entropy decrease is serialized and stored, retaining the mapping relationship with the action parameter index, so as to provide direct input for subsequent global extremum search to lock the optimal action combination; Through the nonlinear transformation processing based on the Shannon entropy formula, the predicted variance map dataset generated in the previous step is transformed into a comparable sequence that quantifies the information gain potential of each candidate action, thereby significantly highlighting the knowledge update value of high variance actions. For example, in the control board of a heater, the candidate set temperature fine-tuning range is set to a value of Nine candidate action points were generated, consisting of temperatures of 0.3℃, 0℃, and +0.3℃, with durations of 2 minutes, 5 minutes, and 8 minutes, respectively. The prediction variance values ​​obtained through step S6.1 were 0.05, 0.02, 0.15, 0.08, 0.03, 0.12, 0.20, 0.18, and 0.09, respectively. Normalization was performed on each prediction variance value, and a probability quantization factor was calculated. =Variance value / Sum of variance values, the sum is 0.92. The probability quantization factors are obtained using methods such as 0.05 / 0.92≈0.054, and these are then... Substitute the values ​​into the Shannon entropy formula to calculate the entropy value, for example, the action point ( The entropy value corresponding to 0.3℃, 2min is 0.054·log(0.054)≈0.155. After calculating the entropy value for all action points, the predicted variance decrease is combined with the expected decrease. For example, the predicted variance decrease for the action point (0.3℃, 2min) after the model update is approximately 0.07. The expected posterior entropy decrease is obtained by multiplying the entropy value by the decrease. In this embodiment, the predicted variance for the action point (+0.3℃, 8min) is 0.20. ≈0.217, entropy value ≈ 0.217·log(0.217) ≈ 0.342, the decrease in prediction variance is 0.10, and the expected decrease in posterior entropy is approximately 0.0342, which is significantly higher than other action points among all candidates, indicating that this configuration has the highest information gain potential in active sampling. After this step is completed, the output is a sequence of expected decreases in posterior entropy with action parameter indices, providing accurate input for the global extreme value search in S6.3, and improving the targeting and effectiveness of the final control action selection; S6.3: Perform a global extreme value search algorithm on the expected value sequence of posterior entropy decrease, compare the information gain potential values ​​of all candidate action points, and select the optimal action index that can maximize the posterior entropy decrease, so as to lock the target control action parameter combination with the most information gain. Based on the numerical distribution of the expected value sequence of posterior entropy decrease, a global search traversal range index matrix is ​​constructed, and the combination of the set temperature fine-tuning amplitude and maintenance duration of each candidate action point is mapped as a unique index label and a bidirectional correlation is established with the corresponding expected value. A full-domain scan operation is performed on the index matrix, and a global extreme value search algorithm is called to compare all elements of the expected value sequence. The comparison operator is used to identify the position index of the largest value in the current dataset. A threshold filtering mechanism is introduced into the global extreme value search algorithm. By setting a lower limit value of the expected information gain, action points that do not meet the minimum information update potential are eliminated, ensuring that the search results are concentrated in the high potential region. The initial maximum value index obtained through global extreme value search is compared and analyzed locally with other high value indices in the neighborhood. The local density evaluation method is used to verify whether the action corresponding to the index has stability and uniqueness in the parameter space. Based on the verified maximum value index, the optimal action index is locked and marked as the target control action parameter combination index with the highest information gain. By using the above processing method, the expected value sequence of posterior entropy decrease in the previous step is transformed into a unique index that can directly point to the optimal action combination, thereby achieving high precision and high confidence in the control action parameter search process. For example, in the low-dimensional action space of an embedded heater control board, the candidate set temperature fine-tuning range is [-0.5℃, -0.3℃, 0℃, +0.3℃, +0.5℃], and the candidate maintenance duration is [1min, 2min, 3min], forming a total of 15 candidate action points. Based on step S6.2, the expected value sequence of posterior entropy decrease for each action point is calculated, for example, using... This indicates that a global extremum search is performed, expressed by the formula: ,in Given the expected value sequence, the search yields the index 8 corresponding to the maximum value of 0.35. A threshold is then set. After filtering, the index still met the conditions, and local density verification showed that the information gain values ​​of its neighborhood were all less than the specified value, confirming its stability and uniqueness. Finally, the action combination corresponding to index 8 was determined: temperature fine-tuning by +0.3℃, and a duration of 2 minutes. The execution results showed a significant decrease in model prediction variance during subsequent active sampling, a faster update rate of the energy-saving and comfort trade-off curve, and an effective improvement in the adaptive capability of the control strategy. S6.4: Based on the index position of the target control action parameter combination, the specific set temperature fine-tuning amplitude value and maintenance duration value are parsed from the candidate action space, and the legality is verified in combination with the current equipment operating status constraints, so as to generate an active sampling control command to be sent with clear execution parameters. S6.5: Perform protocol encapsulation processing on the active sampling control command to be sent, add a timestamp tag and an active sampling mode identifier to form the final control data packet, so as to output the most information-gain active sampling control command that can directly drive the heater to perform temperature fine-tuning and duration maintenance operations.

[0017] Step S7: Execute the active sampling control command to drive the heater to adjust its operating state, and capture in real time whether the user generates new operational feedback and corresponding environmental disturbance data under the new operating state, so as to generate a closed-loop verification dataset containing execution results and new feedback information. Specifically, this includes: S7.1: Based on the set temperature fine-tuning amplitude and maintenance duration parameters contained in the active sampling control command, pulse width modulation mapping is performed on the output power duty cycle and fan speed level of the heater unit to generate the underlying hardware control signal sequence that drives the actuator. S7.2: The underlying hardware control signal sequence is used to drive the heater into a preset fine-tuning operation state, and the surface temperature change slope of the heater, the ambient temperature fluctuation curve and the real-time power consumption data of the device are collected simultaneously to generate a multi-dimensional environmental disturbance feature vector characterizing the physical response of this active sampling process. S7.3: Continuously monitor the event trigger log of the human-computer interaction interface within the maintenance duration window, identify whether the user generates new operation feedback behaviors such as manually raising the set temperature, manually lowering the set temperature, or actively shutting down and restarting the device, so as to generate a set of user behavior response events marked with specific timestamps; S7.4: Based on the multidimensional environmental disturbance feature vector and the user behavior response event set, calculate the user's thermal comfort tolerance score and energy consumption change difference under the current fine-tuning operation state, so as to generate an instant performance evaluation index that quantifies the effect of this active sampling. Based on the multidimensional environmental disturbance feature vector and the user behavior response event set, the feature parsing module is called to extract the numerical features of the ambient temperature fluctuation amplitude, the average slope of surface temperature change, and the change in device power consumption, and they are aligned by timestamp to construct a physical response parameter mapping table. The physical response parameter mapping table is input into the thermal comfort tolerance calculation unit. The difference between each parameter and the mean vector and covariance matrix in the dynamic thermal preference fingerprint distribution is used as the deviation input to calculate the degree of fit between the user's actual response and the expected comfort range. To characterize the impact of temperature deviation on comfort tolerance scores, a Euclidean distance metric is introduced: in, For temperature deviation distance, The current ambient temperature. This is the central value for comfortable temperature; Based on the temperature deviation distance and power change value, the median value of the thermal comfort tolerance score is obtained by weighting according to the tolerance decay function, and then normalized and mapped to the standard score range. The energy consumption change difference calculation unit is invoked to perform an integral calculation on the cumulative power consumption within the maintenance duration window, and the difference is calculated between the integral and the baseline power consumption value. in, Output the difference. The power consumption integral value for the current maintenance duration. Baseline power consumption; The normalized thermal comfort tolerance score and the energy consumption change difference are combined and encapsulated into an instant performance evaluation index structure, and the corresponding active sampling number is marked for correlation and updating. By constructing the above-mentioned real-time performance evaluation indicators, the results of the previous step are transformed into dual indicator data that quantifies user comfort and energy consumption changes, thereby achieving an objective evaluation of the effectiveness of this active sampling and preparing input for subsequent model optimization. For example, during active sampling of a certain heater, the temperature fine-tuning increment was set to 0.3℃, and the duration was 120 seconds. The ambient temperature was observed to gradually decrease from 22.8℃ to 22.4℃, with a surface temperature change slope of -0.05℃ / s, and power consumption decreased from 850W to 820W. The central value μ of the dynamic thermal preference fingerprint was 22.5℃, and the trace value of the tolerance elasticity covariance matrix was 0.09. The temperature deviation distance was calculated as follows: The obtained value is 0.1, and the weighted score after the tolerance decay function is 0.92. The power consumption change difference ΔE, calculated by integration, is -0.03kWh. The negative difference compared with the baseline power consumption indicates the energy-saving effect. In the final output real-time performance evaluation index structure, the comfort tolerance score is 0.92 and the energy consumption change difference is -0.03kWh. In multiple tests, this strategy showed that the comfort tolerance score was stably maintained above 0.9 and the energy consumption change difference tended to be negative, significantly improving the dynamic balance between system comfort and energy-saving performance. S7.5: The real-time performance evaluation index, the corresponding multi-dimensional environmental disturbance feature vector, and the user behavior response event set are spatiotemporally aligned and encapsulated to generate a closed-loop verification dataset containing execution results and new feedback information, which is used to support the subsequent agent model's inducement point update.

[0018] Step S8: Update the induction point position of the sparse variational Gaussian process surrogate model and the statistical parameters of the dynamic hot preference fingerprint distribution using the closed-loop verification dataset, correct the model's cognitive bias towards the user's true preferences, and generate an adaptive optimization control strategy for the next control cycle. Specifically, this includes: S8.1: Perform time-series alignment and outlier removal on the new feedback information and environmental disturbance data in the closed-loop validation dataset to generate a standardized incremental observation sample set as the baseline input data for model update; S8.2: Calculate the Mahalanobis distance between the standardized incremental observation sample set and the existing induced points in the current sparse variational Gaussian process surrogate model, so as to screen out a new set of candidate induced points representing the maximum information gain; S8.3: Utilize the newly added candidate induced point set to execute the induced point relocation algorithm under variational inference, and iteratively correct the original induced point position coordinates and covariance matrix to generate an updated sparse variational Gaussian process surrogate model with higher fitting accuracy. S8.4: Based on the user operation response characteristics implicit in the standardized incremental observation sample set, the mean vector and covariance matrix of the dynamic thermal preference fingerprint distribution are updated online using the recursive Bayesian estimation method to generate the evolved dynamic thermal preference fingerprint distribution that reflects the latest user comfort boundary. S8.5: The updated sparse variational Gaussian process proxy model is fused with the evolved dynamic thermal preference fingerprint distribution to reconstruct the joint response surface of the energy-saving target and the comfort target, so as to generate an adaptive optimization control strategy for the next control cycle with adaptive correction capability.

[0019] The present invention also provides a heater control board, which uses the above-mentioned multi-objective optimization energy-saving control method for heaters to perform multi-objective optimization energy-saving control of heaters.

[0020] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0021] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-objective optimization energy-saving control method for a warm air blower, characterized in that, Includes the following steps: S1: Acquire key user-triggered operation events and synchronized environmental status data during the operation of the heater, and generate raw behavior log sequences; S2: Based on the original behavior log sequence, construct a timestamp-weighted behavior response sequence for each type of key operation event, extract the device power decrease rate, the initial rate of ambient temperature drop, and the time interval between the user's subsequent increase operation within a preset time period after manually lowering the set temperature as joint features, and generate a thermal regulation inertia vector through sliding window aggregation processing. S3: For multiple thermal regulation inertia vectors under different ambient temperature ranges, identify the typical comfort response patterns of individual users in each ambient temperature range and generate a discrete comfort response distribution set; S4: Based on the discrete comfort response distribution set, perform incremental Gaussian mixture modeling on the historical local cluster centers, update the mean vector representing the current comfort temperature center and the covariance matrix reflecting the user's tolerance elasticity to temperature fluctuations, and generate a dynamic thermal preference fingerprint distribution that evolves over time. S5: The dynamic thermal preference fingerprint distribution is injected into the sparse variational Gaussian process surrogate model as prior knowledge. The joint response surface of energy consumption reduction per unit time and fingerprint matching score is fitted using no more than a preset number of induction points to generate an energy-saving-comfort trade-off prediction model. S6: Based on the prediction variance graph of the energy-saving-comfort trade-off prediction model, calculate the next control action parameter that maximizes the decrease in posterior entropy, and generate an active sampling control command.

2. The multi-objective optimization energy-saving control method for a heater according to claim 1, characterized in that, Following S6, the following is also included: S7: Execute the active sampling control command to drive the heater to adjust its operating state, and capture in real time whether the user generates new operation feedback and corresponding environmental disturbance data in the new operating state, and generate a closed-loop verification dataset; S8: Use the closed-loop verification dataset to update the induction point position of the sparse variational Gaussian process surrogate model and the statistical parameters of the dynamic hot preference fingerprint distribution, correct the model's cognitive bias of the user's real preferences, and generate an adaptive optimization control strategy for the next control cycle.

3. The multi-objective optimization energy-saving control method for a heater according to claim 1, characterized in that, The key operational events include manually increasing the set temperature, manually decreasing the set temperature, actively shutting down and restarting the equipment, long-pressing the fan speed setting, and timeout on the temperature control interface. The synchronized environmental status data includes the ambient temperature at the corresponding moment, the current power of the equipment, the running time, and the slope of the heater surface temperature change.

4. The multi-objective optimization energy-saving control method for a heater according to claim 1, characterized in that, Step S3 specifically includes: Based on the ambient temperature component in the thermal regulation inertia vector, interval mapping and data splitting are performed to divide the continuous ambient temperature values ​​into multiple preset temperature threshold intervals, and the thermal regulation inertia vectors falling into the same temperature threshold interval are aggregated into independent environmental context data subsets to generate a vector group to be clustered. Calculate the Euclidean distance matrix between each thermally regulated inertia vector in the vector group to be clustered and all other thermally regulated inertia vectors, and convert the Euclidean distance matrix into a local density index based on the Gaussian kernel function to generate a local density value sequence. Based on the local density value sequence, the minimum distance from each thermally regulated inertial vector to its neighbor with a higher local density value is calculated as the cutoff distance index. The local density value and the cutoff distance index are then multiplied to generate a decision score. A set of candidate cluster centers with decision scores exceeding a preset threshold is selected. The clustering assignment process is initialized using the candidate cluster center point set. Each remaining thermally regulated inertia vector is assigned to the cluster category of its nearest neighbor with a higher local density value until all thermally regulated inertia vectors have been assigned a category, generating a preliminary clustering result set. Extract the geometric centroid coordinates of all thermal regulation inertia vectors within each cluster of the preliminary clustering result set as local cluster centers, and encapsulate the local cluster centers corresponding to all temperature threshold intervals into structured data objects to generate a discrete comfort response distribution set.

5. The multi-objective optimization energy-saving control method for a heater according to claim 4, characterized in that, The Euclidean distance matrix calculation is specifically performed as follows: for each target thermally regulated inertial vector, the Euclidean distance calculation module is called to accumulate the corresponding feature components of the vector with all other thermally regulated inertial vectors in the same interval according to the element difference sum of squares structure, and a square root operation is applied to the sum of squares to generate a complete Euclidean distance matrix. Each element of the Euclidean distance matrix corresponds to a set of distance values ​​between samples.

6. The multi-objective optimization energy-saving control method for a heater according to claim 1, characterized in that, Step S4 specifically includes: Data on multiple local cluster centers in the discrete comfort response distribution set are obtained. Based on the time decay weight function, the historical local cluster centers are preprocessed with weights to generate a weighted historical cluster center sequence. Based on the weighted historical cluster center sequence, perform iterative calculation of Gaussian mixture model parameters to generate an initial set of Gaussian components; Based on the initial Gaussian component set, the mean vector and covariance matrix of the Gaussian mixture model are corrected online to generate an updated Gaussian component parameter set. Based on the updated Gaussian component parameter set, the mean vector of each Gaussian component is extracted as the current comfort temperature center index, and the trace of the covariance matrix is ​​calculated as the temperature fluctuation tolerance elasticity index, generating structured dynamic thermal preference fingerprint distribution data.

7. The multi-objective optimization energy-saving control method for a heater according to claim 6, characterized in that, The initial Gaussian component set represents the probability density of typical user comfort modes under different ambient temperature ranges, and the updated Gaussian component parameter set reflects the user's latest thermal regulation behavior.

8. The multi-objective optimization energy-saving control method for a heater according to claim 1, characterized in that, Step S5 specifically includes: Based on the mean vector and covariance matrix in the dynamic thermal preference fingerprint distribution, an initial induction point candidate set is generated in a low-dimensional operating space, and the initial induction point candidate set is mapped to prior kernel function parameters. The prior kernel function parameters are correlated with the historical operation data in the original behavior log sequence, the Cramer divergence between the likelihood function of the observed data and the variational distribution is calculated, and the variational lower bound objective function value is generated. Based on the variational lower bound objective function value, the gradient ascent optimization strategy is executed to iteratively update the position coordinates and number of the initial candidate set of induced points, eliminate redundant induced points and adjust the distribution density of key induced points on the joint response surface, and generate the optimal set of induced points. The covariance matrix structure of the sparse variational Gaussian process surrogate model is reconstructed using the optimal set of induction points. Multi-task joint modeling is performed with the energy consumption reduction per unit time as the first output dimension and the fingerprint matching score as the second output dimension to generate a joint response surface. The joint response surface is scanned across the entire domain and its features are extracted. The weighted coupling relationship between the predicted energy efficiency and the predicted comfort satisfaction under different control action parameters is analyzed to generate an energy-saving-comfort trade-off prediction model.

9. The multi-objective optimization energy-saving control method for a heater according to claim 8, characterized in that, The joint response surface includes a mean prediction vector and a variance uncertainty matrix.

10. A heater control board, characterized in that: The multi-objective optimization energy-saving control method for the heater described in any one of claims 1-9 is used for multi-objective optimization energy-saving control of the heater.

Images